Skip to main content Link Menu Expand (external link) Document Search Copy Copied

Home page for the Course Web Page for the 2026-27 Edition

** Read the information at the top of the page, then scroll down to see information about each domain. **

Domain Descriptions

DSC Capstone, 2026-27 @ UC San Diego

Overview

Welcome to the capstone program! The capstone program is a two-quarter sequence (Fall 2026 and Winter 2027) in which you will be mentored by a faculty member or industry expert in their domain of expertise. By the end of Quarter 2, you will design and execute a project from that domain in teams. You can see the projects from last year at dsc-capstone.org/showcase-26.

At a high level, here’s how the capstone program is organized:

Enrollment

Enrollment will begin soon. The available domains are not listed on the Schedule of Classes; instead, they are detailed below. Most domains are run by UCSD faculty, but some are run by industry partners (denoted with an Industry Partner badge).

Use the information here to choose the domain you’d like to enroll in. Once you’ve chosen a domain, all you need to do is enroll in the corresponding discussion section for DSC 180A once registration is open, space permitting. Note that you cannot change domains between DSC 180A and DSC 180B.

All of the information here – domain offerings, section times, descriptions, summer tasks, etc. – is subject to change as mentors provide us with more information.

How should I choose a domain?

You should aim to choose a domain that suits your interests and preparation. By clicking the Read more button underneath a domain, you’ll get to learn more about the mentor, their mentoring style, the prerequisites that they’d like their students to have, tasks that they’d like their students to work on over the summer, and their students’ capstone projects in previous years (if any).

✅ Good reasons to choose a domain:

❌ Bad reasons to choose a domain:

Everything you produce for the capstone will have to be public on the internet for the rest of eternity with you and your mentor’s names attached to it – you want your capstone work to be something that you’re proud of and can talk about on job and graduate school applications. Who do you want writing you a recommendation letter?

What happens in DSC 180A?

In addition to meeting with your mentor each week, there will also be methodology instruction delivered by the capstone coordinator and the methodology course staff. However, the majority of this instruction will occur asynchronously, in the form of readings (like this one). This means that you can mostly ignore the lecture and lab times that appear for DSC 180A on the Schedule of Classes. A few of the lecture slots may be used for the capstone coordinator’s office hours or for one-off guest lectures, but we don’t plan to use the majority of the times.

All prerequisites for DSC 180A will be strictly enforced. The prerequisites for DSC 180A can be found here.

Note that since DSC 180A and DSC 180B are both 4-unit courses, you should expect to spend 12 hours a week on capstone-related work each quarter. Plan your class schedule accordingly – try not to take several time-consuming classes alongside the capstone.

Who is overseeing the capstone?

With any questions about the capstone sequence itself, feel free to email Umesh Bellur (ubellur@ucsd.edu) for now.

With any questions about the content of a particular domain, contact the mentor. With any questions about enrollment, please contact Student Affairs in the VAC.

(back to the outline)

Filter by subject area:

AI/ML Systems
🗣️ Language Models
🧠 Theoretical Foundations
⚙️ Applied Data Science

AI/ML Systems

(back to the outline)

BLOCKCHAIN
Rajesh Gupta • rgupta@ucsd.edu
TA: TBA
D18 4 seats Wed 10-11


Blockchains provide a platform for developing new distributed programs and workflow that provide for various services. It is particularly suited for services that involved asynchronous collaboration of diverse actors (human or agents) to achieve overall system objectives. Among the key capabilities are verifiability, non-volatility/immutability of various transactions as well enforcements of various dependencies in a provably correct manner. In this capstone project, you will explore one such service, design and implement it using smart contracts on a chosen platform (Solidity/Ethereum, Solana, Hyperledger etc). You may also consider building upon past projects such as those for GymCoin, RealEstate, etc.
Read more
  • About: Rajesh Gupta is a distinguished professor and the founding director of Halicioglu Data Science Institute. He serves as dean of the School of Computing, Information and Data Science at UC San Diego. A distinguished alumnus of IIT Kanpur, he received an MS from UC Berkeley and PhD from Stanford. He is a fellow of the IEEE, ACM and the AAAS. He holds QUALCOMM endowed chair and International Chair at INRIA, Rennes, France.
  • Mentoring Style: My mentoring is to listen your progress and plans on a weekly basis and lead you to think through alternatives.
  • Suggested Prerequisites:
  • Summer Tasks: Please study, review basics of Blockchain. Since the Smart Contract programming ecosystem is evolving, please research and practice with potential development platform for your project. You may look at the past projects for suggestions.
  • Previous Project

Efficient AI: Algorithms and Systems for Foundation Models
Zhijian Liu • zhijian@ucsd.edu
TA: TBA
D20 8 seats Wed 1-2 PM


Foundation models have become spectacularly capable, and spectacularly expensive. A frontier LLM costs millions of dollars to train and run; a vision-language-action model controlling a robot has to respond in tens of milliseconds. This domain is about bridging that gap — through algorithms and systems that make large models efficient enough to actually deploy. We approach the problem from three angles. The first is inference-time acceleration via speculative decoding, where a small "draft" model proposes tokens that a larger target model verifies in parallel; building on our group's recent work on DFlash, we study how to schedule draft sizes adaptively and how to extend drafts past their training context with sliding-window attention. The second is model compression via post-training quantization, where shrinking weights to 4 bits or lower reduces memory footprint dramatically but introduces error that accumulates along long chains of thought; building on our recent work on ParoQuant, we study outlier suppression with learned pairwise rotations and co-design quantization algorithms with their inference kernels. The third is efficient VLA architectures, where we profile modern models such as π0.5 and InternVLA-A1 to identify compute-vs-memory bottlenecks, and study trade-offs between discrete (FAST tokenizer) and continuous (flow-matching) action representations on the LIBERO benchmark. The unifying thread across all three: identify the real bottleneck through careful profiling, then design an algorithmic or systems intervention that moves the Pareto frontier of speed and quality.
Read more
  • About: Zhijian Liu is an assistant professor at HDSI, UC San Diego, leading a research group on efficient AI. He received his Ph.D. and S.M. from MIT (advised by Song Han) and his B.Eng. from Shanghai Jiao Tong University, with a gap year as a research scientist at NVIDIA. His research on efficient algorithms and systems has been adopted across industry. He has been recognized with the Qualcomm Innovation Fellowship and as a Rising Star in ML & Systems (MLCommons) and in Data Science (UChicago, UCSD).
  • Mentoring Style: Capstone students will be integrated into my research group. Each team will be paired with a PhD student as a day-to-day technical contact, but I will personally lead the weekly section every week. I am hands-on in early Quarter 1 to ensure each team has a tractable problem and a working baseline, and progressively more hands-off as students take ownership. I expect students to read papers, debug code, and form opinions independently, but I am very available over Slack/email between meetings. The best students will treat the capstone as a real research project, not a class assignment.
  • Suggested Prerequisites:
  • Summer Tasks: Pick the track that interests you and prepare accordingly: - Speculative decoding: Read the original speculative decoding paper (Leviathan et al., arXiv:2211.17192) and skim EAGLE-3, Medusa, and the DFlash repo (https://github.com/z-lab/dflash). Run a small Qwen model locally with HuggingFace transformers and make sure you understand the KV cache. - Quantization: Read the AWQ paper (Lin et al., arXiv:2306.00978) and the ParoQuant paper (arXiv:2511.10645), and skim the ParoQuant repo (https://github.com/z-lab/paroquant). Try running a quantized 4-bit model locally with vLLM. Make sure you understand why outliers in weights and activations cause large errors at low bit-widths. - Efficient VLA: Read the π0.5 paper (arXiv:2504.16054) and the FAST tokenizer paper (arXiv:2501.09747). Get LeRobot (https://github.com/huggingface/lerobot) running locally on a small policy. Regardless of track, brush up on PyTorch internals (autograd, torch.compile) and profiling tools (torch.profiler, Nsight Systems), and practice tracing a forward pass through an unfamiliar repo. If you want a head start, attempt the baseline measurement for your chosen track (acceptance length on MATH-500, MMLU-Pro accuracy under INT4, or an nsys profile of π0.5) and bring the numbers to our first meeting.
  • Previous Project

Agentic workload analysis on at scale health / bio tech use cases to determine optimzal CPU, GPU and memory architectures.
Renu Raman, Simran Arora, Daniel Kang, Brian Chin • renu.raman@terizza.com, simarora@stanford, b5chin@ucsd.eduedu
TA: TBA
D29 8 seats >2 hours/week Industry Partner


initiative focuses on "Agentic Compute," which involves LLMs interacting with external tools, databases, and environments. This requires benchmarking not just raw token generation, but the entire system's ability to process compound tasks. Defining the Mixed Workload A realistic agentic workload consists of three interleaved components: LLM Reasoning and Planning: The model analyzes the user query and decides which tools to use. SQL Database Lookups (Structured Data): The agent generates SQL queries to retrieve precise tabular data (e.g., patient records, financial transactions). Vector Database Lookups (Unstructured Data): The agent performs semantic searches against a vector database to retrieve relevant documents or context (e.g., research papers, clinical notes). Knowledge Graph Lookups (KG): The agent or the query plan looks up knowledge graphs that has built the entity relationships for large and complex enterprise use cases.
Read more
  • About: www.linkedin.com/in/renuraman, https://arorasimran.com/, https://www.linkedin.com/in/bryan-chin-420806/
  • Mentoring Style: Will be hands on work
  • Suggested Prerequisites:
  • Summer Tasks: Understand Spider2.0, Inference engines (disaggregated) including kernels and execution stages (prefill/decode), industry published benchmarks on agentic coimpute
  • Previous Project

Autonomous, Multimodal Data Ecosystem for Enterprise Intelligence
Renu Raman, Vikram Joshi, Jack Brzeznski • renu.raman@terizza.com, vikram.joshi@terizza.com, jack.brzeznski@terizza.com
TA: TBA
D32 8 seats Flexible. Will work with the students. for 1-2 hours/week Industry Partner


Topics to explore Sub-Project 1: The Compute.AI Microkernel & Infrastructure Focus: Foundational Database and Compute Engine Development Description: This sub-project focuses on building the core engine of the Compute.AI platform. Students will develop a highly concurrent, memory-centric microkernel that supports AI-driven memory tiering and dynamic scheduling The infrastructure will be designed to handle mixed-mode compute, seamlessly integrating relational SQL, vector databases, and graph databases to support massive enterprise workloads. Sub-Project 2: Conversational UI & Code Generation Engine Focus: User Experience and Automated Analytics Description: Bridging the gap between complex data and end-users, this team will design an intuitive conversational dashboard . A key feature will be the integration of an AI-powered code generation system capable of translating natural language queries into executable SQL and Python code . This will allow users to dynamically interact with both relational and multimodal data sources without needing deep technical expertise. Sub-Project 3: Hybrid AI for Tabular and Multimodal Integration Focus: Data Processing and Machine Learning Description: Addressing the "tabular data conundrum," this sub-project will develop hybrid AI models that effectively process structured enterprise data alongside unstructured text and images . By utilizing vector embeddings and retrieval-augmented generation (RAG), the team will ensure that the Compute.AI platform can extract deep insights from diverse data types, optimizing the system for both classical ML and modern deep learning. Sub-Project 4: Agentic AI for Domain-Specific Research (Healthcare Use Case) Focus: Autonomous Application and Discovery Description: Serving as the primary application layer, this project will deploy long-running AI agents on top of the Compute.AI infrastructure to automate complex research tasks Using the medical field as a proof-of-concept, the agents will autonomously ingest literature, query multimodal databases (such as radiology images and clinical text), and generate hypotheses, demonstrating the platform's capability to compress months of research into hours.
Read more
  • About: Renu Raman's 30-year career spans seven ventures across silicon, systems, in-memory computing, databases, and cloud infrastructure, consistently incubating businesses and pioneering technologies like RISC, multi-core/threading, the camera phone, and in-memory accelerated computing, scaling businesses from $0 to $10B. Vikram is a serial entrepreneur and coder, his technical background spans multiple disciplines such as databases, operating systems, parallel and distributed systems, AI/ML, storage, video streaming, and computer graphics. His company ioTurbine specialized in software to accelerate storage using SSDs (acquired by Fusion-io). Xcalar, a relational compute platform, was used by top-tier Wall St. banks and financial institutions. Expertise in Generative AI, ML, orchestration for long-time reasoning, GenAI agentic cybersecurity, physical AI, and Analytics. An educator with 20+ years of experience in consulting, system design, and directing execution for enterprise projects.
  • Mentoring Style: Yes, this will be part of the overall research efforts in collaboration with UCSD
  • Suggested Prerequisites:
  • Summer Tasks: We will provide a collection of papers and github material
  • Previous Project

Efficient LLM Inference Systems: Scaling Large Language Models from Single GPU to Datacenters
Hao Zhang • haz094@ucsd.edu
TA: https://zhongyan0721.github.io/tpu-dflash/
D38 6 seats Monday 2pm


The deployment of large language models (LLMs) is now central to modern AI infrastructure, yet inference remains constrained by GPU memory, latency, throughput, and cost. As models scale to hundreds of billions of parameters and context windows stretch to millions of tokens, naive serving strategies break down. The field has responded with a wave of systems-level innovations: PagedAttention and KV cache management (vLLM, SOSP'23), disaggregated prefill/decode serving (DistServe, OSDI'24), speculative decoding, continuous batching, KV cache offloading to CPU and SSD (Mooncake, LMCache), distributed inference across heterogeneous accelerators, and specialized scheduling for long-context and agentic workloads. This domain explores the open problems in LLM serving — how to push throughput and latency Pareto frontiers further, how to serve mixture-of-experts and reasoning models efficiently, how to handle KV cache at exabyte scale, and how to co-design model architectures with serving systems. Students will gain hands-on experience with vLLM and SGLang, profile real workloads, and propose interventions that move the state of the art. Suggested references: Kwon et al., Efficient Memory Management for Large Language Model Serving with PagedAttention, SOSP 2023 (vLLM) Zhong et al., DistServe: Disaggregating Prefill and Decoding for Goodput-optimized LLM Serving, OSDI 2024
Read more
  • About: Hao Zhang is an Assistant Professor in Halıcıoğlu Data Science Institute and the Department of Computer Science and Engineering at UC San Diego. Before joining UCSD, he was a postdoctoral researcher at UC Berkeley working with Ion Stoica from 2021 to 2023. He completed his Ph.D. in Computer Science at Carnegie Mellon University with Eric Xing from 2014 to 2020. During his PhD, he took a leave to work at the ML platform startup Petuum Inc from 2016 to 2021. His research lies at the intersection of machine learning and systems. His work includes FastVideo, DistServe, vLLM, Chatbot Arena, Vicuna, Alpa, Poseidon, and Petuum. His research has been recognized with a Sloan Research Fellowship (2026), a Google ML and Systems junior faculty award (2025), the Jay Lepreau Best Paper Award at OSDI'21, and an NVIDIA Pioneer Research Award at NeurIPS'17. He was nominated for MIT TR35 (China) in 2024. He joined Snowflake through acquihiring in 2023 and also co-founded the nonprofit LMSYS Org, which maintains many popular open models, evaluation platforms, and systems.
  • Mentoring Style: hands off
  • Suggested Prerequisites:
  • Summer Tasks: N/A learning pytorch / vllm / sglang
  • Previous Project

Language Models

(back to the outline)

Trustworthy Large Language Models
Lily Weng • lwegn@ucsd.edu
TA: TBA
D05 8 seats Monday 4 pm


Developing Methods to Make Language Models More Trustworthy This capstone project focuses on designing and evaluating methods to improve the trustworthiness of large language models (LLMs), with an emphasis on reasoning tasks. While modern LLMs have achieved strong performance in accuracy, they often fall short in interpretability, faithfulness, and reliability, which are key properties required for safe and dependable deployment. The broader goal is to move beyond black-box AI systems toward interpretable, controllable, and reliable language models that can be safely used in high-stakes applications such as decision support, scientific reasoning, and autonomous systems.
Read more
  • About: Lily Weng leads the Trustworthy Machine Learning Lab at UCSD. Her research vision is to make the next generation AI systems and deep learning algorithms more robust, reliable, explainable, trustworthy and safer.
  • Mentoring Style: students are expected to lead the projects
  • Suggested Prerequisites:
  • Summer Tasks: 1. Students should be familiar with Pytorch and have taken DSC 140B or the equivalent courses in deep learning. 2. Students are expected to read one of the papers and successfully set up one of the repos below: (1) https://lilywenglab.github.io/Training_Trustworthy_LRM_with_Refine/ (2) https://lilywenglab.github.io/ReflCtrl/
  • Previous Project

LLM Applications
Jingbo Shang • jshang@ucsd.edu
TA: TBA
D07 4 seats Likely Wednesday morning/noon.


We will explore the LLM applications together. Learn, explore, and brainstorm in Quarter 1. Finalize a project and team up to deliver in Quarter 2. Working apps are expected at the end of Quarter 2. This field is moving fast, so I don't have any preset topic at this moment.
Read more
  • About: I have a broad interests in data mining, natural language processing, and machine learning. My goal of research is to build models without annotating any data.
  • Mentoring Style: Small group discussion with me.
  • Suggested Prerequisites:
  • Summer Tasks: Play with LLM APIs, such as GPT and Claude. Use Coding Agent, such as Claude Code and Codex.
  • Previous Project

Steering and monitoring LLMs
Misha Belkin •
TA: TBA
D08 10 seats Wed 9am


Activation steering and monitoring is a powerful methodology for understanding and controlling behavior in LLMs. We will implement steering methods and address various problems related to data representation in LLMs and modifying their behavior.
Read more
  • About: https://misha.belkin-wang.org/bio.html
  • Mentoring Style: Several groups embarking on potentially related projects.
  • Suggested Prerequisites:
  • Summer Tasks: https://arxiv.org/abs/2502.03708
  • Previous Project

Programming Cognitive Bias in Social Agents
Haojian Jin • haojian@ucsd.edu
TA: TBA
D09 8 seats Either in-person or fully zoom, depends on the students' preference. Prefer to meet at the end of a weekday.


Reproducibility and controllability are fundamental to scientific research. Yet implicit natural language descriptions — the dominant approach for specifying social agent behaviors — often fail to yield consistent behavior across models. CoBRA operationalizes validated social science knowledge as reusable environments for AI — enabling explicit, quantitative control.
Read more
  • About: I am a human-centered system researcher. My research lies at the intersection of human-computer interaction, privacy and security, and systems, focusing primarily on two conflicting questions: How can we design systems that meaningfully improve people’s lives? How can we prevent those systems from becoming unsafe as they grow more proactive and powerful? ?
  • Mentoring Style: Integrate capstone students into our on-going research project.
  • Suggested Prerequisites:
  • Summer Tasks: https://cobra.clawder.ai/ Read the paper and understand the codebase
  • Previous Project

Ensemble of agents
Yoav Freund • yfreund@ucsd.edu
TA: TBA
D10 4 seats 1 hour per week


This project combines two subjects: Ensemble methods and Agent based AI. Boosting [1], Bagging [2] and random forests [3] are popular machine learning methods that combine the output of weaker learning algorithms to create a single highly accurate rule (or model). It is based on Agent based AI is all the rage these days. It refers to having independent LLM sessions that are running in parallel and communicating with each other. The idea is to merge these two directions to construct boosting and bagging systems where a swarm of agents implements the basic learners and a coordinator agent coordinates the swarm to create a combined accurate and reliable classifier. Some work in a similar direction was pushed in [3] [1] Freund, Yoav, and Robert E. Schapire. "A decision-theoretic generalization of on-line learning and an application to boosting." Journal of computer and system sciences 55.1 (1997): 119-139. [2] Breiman, Leo. "Bagging predictors." Machine learning 24.2 (1996): 123-140. [3] Breiman, Leo. "Random forests." Machine learning 45.1 (2001): 5-32. [4] Alafate, Julaiti, and Yoav S. Freund. "Faster boosting with smaller memory." Advances in Neural Information Processing Systems 32 (2019).
Read more
  • About: My work is in the area of Machine Learning. I am known for my work with Rob Schapire on the Adaboost algorithm.
  • Mentoring Style: I plan to assign tasks to students. Tasks can be either to present a paper or to develop a piece of software and integrate it into the system.
  • Suggested Prerequisites:
  • Summer Tasks: Read the papers in the description. Educate yourself on Cursor and on claude agents SDK
  • Previous Project

Steering of diffusion models
Yusu Wang • yusuwang@ucsd.edu
TA: TBA
D12 12 seats either Tue or Wed morning at 9am.


Diffusion models are a family of popular generative models for images, and increasingly, for language and graph data as well. However often we wish to have certain control over what is being generated and be able to steer the output towards certain types, concepts, or satisfying some constraints. Recent work (e.g., our work here: https://arxiv.org/abs/2602.11395) has shown that activation steering can be quite effective both for languages and images. This capstone project will explore the latent space of diffusion models across "time" as well as improved steering strategies of diffusion based generative models to further improve their efficiency, accuracy, and broader applicability.
Read more
  • About: Graph representation learning, diffusion models, neural algorithmic reasoning
  • Mentoring Style: One of my postdocs who has great experiences with steering diffusion models (both hands-on and on the theoretical front) will co-lead this project with me. He will also be able to serve as a TA for this capstone.
  • Suggested Prerequisites:
  • Summer Tasks: (1) https://arxiv.org/abs/2206.00364 see github here: https://github.com/nvlabs/edm (2) https://arxiv.org/abs/2602.11395
  • Previous Project

Community-Centered Discrimination Audits of LLMs - Bias Rapid Action Teams
Stuart Geiger •
TA: TBA
D15 6 seats Wed 10-11am


This capstone will work with community members to audit pretrained Large Language Models for discrimination and bias, using perturbation-based or controlled-experimental methods. These systematically vary a template prompt along a potential type of discrimination, then observe differences in outputs. For example, if you ask ChatGPT (or TritonGPT) to act as a college admissions reviewer, does an application's score change if it references the Mens vs Womens basketball team? Or being on the lacrosse versus basketball team? Or being from La Jolla versus San Ysidro? These methods are relatively simple from a statistics perspective, but the hard part is knowing what kinds of discrimination are of most concern to the people who will be impacted by model outputs and creating real-world template prompts that test for those concerns. This capstone will be centered around **talking and listening to real people** about their concerns with LLMs in real-world contexts, then using our data science expertise in a more consulting-style mode. If a team chooses university admissions, they might work with students, high school counselors, professors, and/or admissions staff. All students must take and pass the 3-hour UCSD/CITI IRB Human Subject Protection Training online course (Social and Behavioral Basic) by week 3 of Fall.
Read more
  • About: I’m a social scientist with a background in the humanities, especially history and philosophy of science and technology, but I have enough expertise in computer science and data science to make trouble. I believe that data science systems should be fair, transparent, and accountable to the public, but that most are currently not. A lot of my research is in community-centered content moderation NLP systems for user-generated content, especially Wikipedia, where I formerly worked on their ML models and systems.
  • Mentoring Style: I will be the point of contact and there every week, but may bring in collaborators and my grad student advisees. I intentionally do not run a "lab", but I do have a "constellation of collaboration." Students can choose their own particular context in which LLMs are deployed and which kinds of community members / impacted people they want to consult.
  • Suggested Prerequisites:
  • Summer Tasks: Get familiar with querying LLM APIs that use an OpenAI Chat Completions compatible API, try out openrouter.ai. Read an example of a perturbation-based audit study: - https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0318500 For more readings, see - https://auditlab.stuartgeiger.com Take UCSD/CITI IRB Human Subject Protection Training online course (Social and Behavioral Basic), must complete by week 3 of Fall, but good to do it earlier. Register at citiprogram.org and see this video for how to register: - https://www.youtube.com/watch?v=hOAgfK93QXg
  • Previous Project

Steering Large Language Models
Tianhao Wang • tianhaowang@ucsd.edu
TA: TBA
D19 8 seats Friday


Large language models are increasingly used as general-purpose assistants, but they are still difficult to control reliably: they may answer when they should abstain, express unjustified confidence, refuse benign questions, or produce reasoning traces that appear coherent while masking shallow shortcuts. This capstone domain focuses on LLM steering for reliability and reasoning, especially methods that make models better calibrated about what they know, when they should say “I don’t know,” and how to preserve or improve reasoning ability while changing model behavior. Students will explore the broader problem space of steering language models through prompting, supervised fine-tuning, preference optimization/RLHF-style methods, representation or activation steering, uncertainty estimation, and selective prediction/abstention. A central tension is that interventions that improve one behavior, such as safer refusal or higher abstention, can unintentionally degrade another, such as mathematical reasoning, helpfulness, or generalization to out-of-distribution tasks. Students may investigate questions such as: how do we measure whether a model abstains for the right reasons; how do chain-of-thought or reasoning traces interact with confidence and refusal; can we steer internal representations toward honesty or uncertainty without suppressing useful reasoning; and how can synthetic or benchmark tasks reveal shortcut learning?
Read more
  • About: I am an Assistant Professor in the Halıcıoğlu Data Science Institute at University of California, San Diego. I am broadly interested in various aspects of machine learning, optimization, and statistics. Prior to UCSD, I was a Research Assistant Professor in the Toyota Technological Institute at Chicago from 2024 to 2025. Before that, I received my Ph.D. from the Department of Statistics and Data Science at Yale University.
  • Mentoring Style: I plan to run this capstone as a small research group with weekly meetings. Early on, I will be more hands-on in helping students understand the area, choose papers to read, and design feasible replication experiments. Over time, students will take more ownership of their projects, while I provide regular feedback on research questions, experimental design, implementation issues, and interpretation of results. Students should expect to work independently between meetings and come prepared to discuss progress and challenges.
  • Suggested Prerequisites:
  • Summer Tasks: Students should use the summer to build basic background in LLM reliability, steering/alignment, calibration, and abstention. It would also be helpful to review core machine learning concepts and get comfortable running small experiments with Python, PyTorch, and Hugging Face Transformers. In Quarter 1, students will likely start with a small replication project related to LLM abstention, calibration, refusal behavior, or reasoning/steering, so students are encouraged to think broadly about which part of this space they find most interesting.
  • Previous Project

Empirical Study of LLM and Coding Agent Best Practices
Armin Schwartzman <armins@ucsd.edu>, Justin Eldridge <jeldridge@ucsd.edu> •
TA: TBA
D23 16 seats Wednesdays 3 pm (with flexibility around it if needed)


Practitioners have accumulated a large body of advice on how to get the most out of large language models and coding agents, such as writing a plan before coding, using test-driven development, using structured prompts, decomposing tasks across sub-agents, and so on. Some of this advice traces back to published research, but much of it circulates as rules of thumb whose effect sizes are unknown or which have never been re-validated on current models or in agentic coding settings. This project asks which practices actually hold up under empirical scrutiny. Students will survey the relevant literature to identify the most common “tricks” employed in practice, translate these into concrete experimental conditions, and design controlled studies. Specific research questions will be developed with the team. Students will gain hands-on experience with experimental design, statistical analysis of stochastic AI systems, and the modern coding-agent ecosystem — and, ideally, produce findings that contribute to a more rigorous understanding of how to use these tools well.
Read more
  • About: Armin Schwartzman: With an undergraduate degree in electrical engineering, I discovered statistics for my PhD and have been doing data science since then (even when it wasn't called by that name). Much of my work involves signal and image analysis, but I'm interested in many theoretical and applied problems, even philosophical. Outside of academia, I like doing music, dancing, swimming, surfing, and more. Justin Eldridge: Justin Eldridge is an Associate Teaching Professor in HDSI. He obtained his PhD in computer science at The Ohio State University as a Presidential Fellow, along with BS degrees in physics and applied math. His research focus lies in statistical machine learning theory, with an emphasis on unsupervised learning and clustering in particular. His research while a PhD student received the best student paper award at COLT 2015 and a full oral presentation at NeurIPS 2016. Justin joined HDSI in 2018, where he develops and teaches courses in both the theoretical and practical foundations of data science and machine learning.
  • Mentoring Style: The capstone will be managed in two parts. In the first quarter, Armin Schwartzman will work with students on the basics of experimental design and statistical analysis, illustrated by experiments on small LLMs. In the second quarter, Justin Eldridge will guide students on larger LLM experiments and their consequences in coding, teaching, psychology, etc. Students are expected to take ownership over the project. This implies taking initiative in learning about the topic (from the assigned material and other sources), implementing the methods in code, being resourceful when needing help, and asking questions. Students are expected to put in their best effort, plan their time over each quarter, make substantial progress each week, report on it each week, and come up with an action plan for the next steps (as opposed to waiting for the mentor to give instructions). In other words, be independent and ask for help when needed.
  • Suggested Prerequisites:
  • Summer Tasks: Practice using LLMs
  • Previous Project

Epistemic Engines: Turning Raw Data into Living Knowledge Bases for AI Agents
Ryan Lingo and Rajeev Chhajer • ryan_lingo@honda-ri.com and rajeev_chhajer@honda-ri.com
TA: TBA
D25 12 seats Weekly at the student team discretion Industry Partner


Data is everywhere, but most of it sits in forms that are hard to use: buried in PDFs, scattered across spreadsheets, locked in databases, streaming from sensors, hiding in web pages. What if we could take all of that, normalize it into a clean, readable format like Markdown, and hand it to AI agents that build and maintain structured knowledge from it? That's the core question behind this domain. We'll explore how to turn messy, heterogeneous data (structured and unstructured) into a common substrate that LLM-powered agents can read, reason over, link together, and keep current. From there, the question opens up: what does it mean to "create knowledge" from data? That could look like an agent that synthesizes a living wiki from research papers, one that tracks and reconciles contradictory claims across news sources, one that monitors a domain and surfaces what's missing, or something else entirely. A hard problem running through all of this is evaluation: how do you measure whether a system is actually producing good knowledge? Factual accuracy is just the starting point. You also need to think about coverage, coherence, source grounding, staleness, and whether the system knows what it doesn't know. Designing the right evaluation framework may be as interesting as building the system itself. A key starting point is Andrej Karpathy's "LLM Wiki" pattern (https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f), where raw sources flow into a folder and an LLM agent compiles and maintains an interlinked Markdown knowledge base, browsable in Obsidian. Students will build on this idea and push it in their own direction. Relevant references include Lewis et al. on retrieval-augmented generation (https://arxiv.org/abs/2005.11401), Hogan et al. on knowledge graphs (https://arxiv.org/abs/2003.02320), Gärdenfors' Knowledge in Flux on belief revision, and the Stanford STORM project on LLM-driven research synthesis (https://storm.genie.stanford.edu/).
Read more
  • About: Rajeev Chhajer is the Chief Engineer at Honda Research Institute USA and leads the Software-defined Intelligence team at 99P Labs. He is a founding member of 99P Labs, a research initiative dedicated to developing sustainable technologies and innovative approaches to global challenges. His research focuses on smart city ecosystems, embedded systems, and connectivity to support sustainable and efficient mobility. Ryan Lingo is an Applied AI Research Engineer and Developer Advocate at 99P Labs. His work focuses on intelligent systems, with research interests in large language models, synthetic data, and applied machine learning. He has an academic background in philosophy and has held roles in data science and software engineering, with experience spanning academic research, industry, consulting, and early-stage startups.
  • Mentoring Style: We plan to take an engaged but student-led approach to mentoring. We'll work closely with the students throughout the project, meeting regularly, providing guidance, and being available for feedback and support. That said, we're looking for high-agency students who are excited to take ownership of their learning and direction. The best mental model for this capstone is "learning in public." Students will play an active role in shaping the plan and setting objectives. Rather than being given step-by-step instructions, they'll be encouraged to explore, make decisions, and figure out how to execute their ideas, with our mentorship to guide the way. We'll help them think critically, problem-solve, and communicate their process and outcomes clearly. While we won't dictate tasks at a granular level, we'll be present every week and ensure they have the support and structure they need to succeed.
  • Suggested Prerequisites:
  • Summer Tasks: 1. Build Karpathy's LLM Wiki yourself. This is the most important summer task. Follow Karpathy's gist (https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f) to set up an Obsidian vault with a raw/ folder for sources and a wiki/ folder where an LLM agent maintains interlinked Markdown notes. Use OpenCode (https://opencode.ai) as your agentic coding tool. It's free, open-source, and works with 75+ LLM providers including local models via Ollama. Drop in 10-20 sources on a topic you care about (research papers, articles, transcripts) and have the agent compile a wiki from them. Browse the result in Obsidian's graph view. The goal is to feel both the power and the limitations: where does the synthesis break down? What gets lost? What would you add? 2. Read these: Karpathy's original thread on LLM knowledge bases: https://x.com/karpathy/status/2039805659525644595 Lewis et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" (2020): https://arxiv.org/abs/2005.11401 Hogan et al., "Knowledge Graphs" (2021): https://arxiv.org/abs/2003.02320 Skim the Stanford STORM project: https://storm.genie.stanford.edu/ 3. Get familiar with the tools. Install and try OpenCode (https://opencode.ai/docs). Install Obsidian (https://obsidian.md) and explore plugins like Dataview and Graph View. If you haven't used a vector database (Chroma, Weaviate, etc.), set one up and run some queries. If you've never worked with graph databases (Neo4j), watch an intro tutorial. 4. Think about what interests you. The domain is deliberately open. Some directions you might explore: evidence grounding (can every claim trace back to a source?), belief revision (what happens when new evidence contradicts old beliefs?), contradiction detection across sources, confidence scoring, hypothesis generation (what's missing from the knowledge base?), cross-domain transfer (can patterns from one field unlock insights in another?). Come to fall with at least one domain you'd want to point this system at (scientific literature, journalism, clinical data, competitive intelligence, something else) and a rough sense of which problem within it excites you most. 5. Start thinking about evaluation. This is one of the hardest parts of the project and worth thinking about early. If your system produces a knowledge base, how do you know it's any good? You might evaluate factual accuracy against ground truth, measure coverage relative to a source corpus, check whether every claim is grounded in evidence, test how well the system handles contradictions, or assess whether it correctly identifies gaps. There's no single right answer here. Part of the research is figuring out what "good knowledge" means for your specific domain and building a way to measure it. Come to fall with some initial thoughts on what you'd want to measure and how. Technologies you'll use: Python, OpenCode, Obsidian, LLM APIs, Markdown as a knowledge substrate, possibly vector databases and graph databases. If any of these are unfamiliar, spend time with them over the summer.
  • Previous Project

Red-Teaming the Swarm: Autonomous Auditing for Financial LLMs
Dusan Bosnjakovic (dusan_bosnjakovic@intuit.com); Perez Ogayo (perez_ogayo@intuit.com); Dhruvi Kothari (dhruvi_kothari@intuit.com); Victor Calderon (victor_calderon@intuit.com) • dusan_bosnjakovic@intuit.com; perez_ogayo@intuit.com; dhruvi_kothari@intuit.com; victor_calderon@intuit.com
TA: TBA
D30 5 seats Thursdays - 2-3 pm (but can be modified) Industry Partner


As financial institutions transition from single-turn LLM queries to complex agentic workflows, the need for rigorous, domain-specific evaluation has become critical. This project focuses on building a modular benchmarking framework designed to stress-test vanilla, fine-tuned, and agentic systems for performance, safety, and reliability. Students will evaluate technical hardware metrics—such as Time to First Token (TTFT) and tokens per second—alongside qualitative assessments of accuracy, hallucination, and bias. A central research focus will be the phenomenon of multi-turn drift, where agentic performance or safety alignment degrades during long-horizon reasoning, recursive workflows, or memory retrieval. The core objective is to develop an extensible architecture that allows users to overlay their own models, agentic configurations, and custom risk-tolerance thresholds. Beyond simply running tests, students will be expected to maintain a research-level of scientific rigor, applying statistical methods to prove the validity and reliability of their benchmarks. This includes utilizing cutting-edge industry standards while developing novel, domain-grounded frameworks for high-stakes financial tasks like tax law interpretation or credit risk assessment. By the end of the sequence, students will have engineered a production-ready, peer-reviewed infrastructure for the autonomous auditing of AI systems, ensuring they remain robust against both adversarial "red-team" attacks and the inherent instabilities of multi-agent collaboration.
Read more
  • About: Bio 1: Dusan Bosnjakovic: Dušan Bošnjaković is an AI Group Manager at Intuit, specializing in Agentic AI development and enterprise-scale LLMOps. With a strong background in deep technical IP and platform-scale engineering, he focuses on architecting production-grade Generative AI platforms and orchestrating AI governance across the enterprise. Prior to his current role, Dušan served as Director of Data Science at TeleSign and architected real-time machine learning underwriting platforms during his earlier tenure at Intuit. He holds a B.A. and post-graduate research experience in Cognitive Psychology and Psycholinguistics from the University of California, San Diego; Bio 2: Perez Ogayo: Perez Ogayo is a Machine Learning Engineer at Intuit, specializing in LLMOps, agentic workflows, and chatbot development. A published NLP researcher with a focus on low-resource African languages, she focuses on building and productionizing robust Retrieval-Augmented Generation (RAG) systems and Generative AI data pipelines. Prior to Intuit, Perez worked as a Software Engineer at Oracle and an NLP Researcher at Masakhane and Carnegie Mellon University, where she designed and evaluated transformer-based translation models. She holds a Master's degree in Language Technologies from Carnegie Mellon University.; Bio 3: Dhruvi Kothari: Dhruvi Kothari is an AI Data Scientist at Intuit, specializing in Generative AI, text extraction, and synthetic data. A transportation planner-turned-data scientist, she focuses on applying data and design thinking to develop impactful, 0-to-1 AI-native products. Prior to Intuit, Dhruvi worked as a Senior Data Scientist at Urban Footprint and a Transportation Engineer at Fehr & Peers, where she built machine learning models and data pipelines for urban and climate analytics. She holds a Master's degree in Information Data Science from the University of California, Berkeley; Bio 4: Victor Calderon: Victor Calderon is a Machine Learning Engineer at Intuit, specializing in Generative AI and LLMs. An astrophysicist-turned-data-scientist, he focuses on applying Generative AI solutions to customer-related problems. Prior to Intuit, Victor developed and deployed computer vision models and MLOps pipelines at 5x5 Technologies Inc. He holds a Ph.D. in Physics with a focus on computational astrophysics from Vanderbilt University;
  • Mentoring Style: We plan to take an engaged but student-led approach to mentoring. We’ll work closely with the students throughout the project – meeting regularly, providing guidance, and being available for feedback and support. We’re also looking for students that can take ownership of their learning and direction, and can execute on the feedback provided. We’ll help them think critically, problem-solve, and communicate their process and outcomes clearly. Outside of the set office hours, we will do our best to respond to any inquiries within 48 hours, ideally sooner.
  • Suggested Prerequisites:
  • Summer Tasks: We will offer context and background as summer reading. Good to review: Python Pandas PyTorch Multi-Agent architecture LLM Evaluations & Benchmarking
  • Previous Project

Steering and Monitoring LLMs for Safety
Dmitriy Drusvyatskiy • ddrusvyatskiy@ucsd.edu
TA: TBA
D36 10 seats Wednesday mornings.


Large language models are increasingly used in high-stakes settings, but their behavior can still be unreliable, easy to manipulate, or difficult to understand. This group will study methods for steering LLMs toward safer behavior and monitoring them for failures such as jailbreaks, harmful outputs, deception, sycophancy, or over-refusal. We will explore techniques including activation steering, representation engineering, probing, red-teaming, safety classifiers, and automated monitors. Students will read and reproduce recent work on LLM safety interventions and evaluations, then design empirical projects testing how well these methods work across prompts, models, and adversarial settings.
Read more
  • About: Dmitriy Drusvyatskiy received his PhD from Cornell University in 2013, followed by a post doctoral appointment at University of Waterloo, 2013-2014. He joined the Mathematics department at University of Washington as an Assistant Professor in 2014 and was promoted to Full Professor in 2022. Since 2025, Dmitriy is a Professor at the Halıcıoğlu Data Science Institute (HDSI) at UC San Diego. Dmitriy's research broadly focuses on designing and analyzing algorithms for large-scale optimization problems, primarily motivated by applications in data science.
  • Mentoring Style: This will be separate from my research group
  • Suggested Prerequisites:
  • Summer Tasks: Read material on Mechanistic Interpretability, which is readily found online.
  • Previous Project

Baby Language Models and Cognitive Modeling
Alex Warstadt • awarstadt@ucsd.edu
TA: https://lasyayadlapati.github.io/sleep-consolidated-learning/
D40 6 seats Wednesday 11-12


Large Language Models have an impressive ability to learn and use human language, but humans are still the state-of-the-art when it comes to learning language efficiently. We acquire language from 100 million words or less, whereas LLMs are now trained on 10s or *trillions* of words. The BabyLM Challenge (https://babylm.github.io/) is a competition centered around training small "BabyLMs" under constraints inspired by human language learning. The goal of a BabyLM submission is to train a model that learns language as data-efficiently as a human or that simulates properties of human learning and linguistic performance.
Read more
  • About: Alex Warstadt is an Assistant Professor at UCSD with joint appointments in HDSI and the Department of Linguistics. He received his PhD in linguistics in 2022 from NYU under Samuel Bowman, and completed a postdoc in 2024 with Ryan Cotterell. Alex runs UCSD's Learning Meaning and Natural Language Lab (LeMN Lab) which is an interdisciplinary group that uses insights from linguistics to advance and interpret language models and uses advances in machine learning to answer scientific questions about the nature of language.
  • Mentoring Style: We hold weekly meetings with all capstone projects. Groups working on different projects will still be present to learn about and provide input on other projects. PhD students from my lab are invited to provide additional mentorship.
  • Suggested Prerequisites:
  • Summer Tasks: NLP, linguistics, and cognitive science are helpful but not required. Replicate prior BabyLM studies, including training LMs from scratch and evaluating, launching jobs on a UCSD supercomputer cluster such as Nautilus, and tracking training progress and sweeps over several days or weeks using weights and biases.
  • Previous Project

🧠 Theoretical Foundations

(back to the outline)

Development of Sequential Hypothesis Testing Algorithms
Jun-Kun Wang • jkw005@ucsd.edu
TA: TBA
D04 4 seats Tuesdays 4PM


You will read a couple of papers in top AI venues regarding sequential hypothesis testing, which has many applications in checking the validity and reliability of real-use systems, e.g., AI safety & monitoring. You will understand the algorithmic foundations of sequential hypothesis testing and be able to replicate some results in the literature. After mastering the basics, you will try to design your own method, find an interesting new domain that can benefit from sequential hypothesis testing, or do both.
Read more
  • About: I am an assistant professor at HDSI and ECE. My research is centered around optimization and its connections with statistics and machine learning.
  • Mentoring Style: Students will be expected to use Python to implement their algorithms (and should be able to code).
  • Suggested Prerequisites:
  • Summer Tasks: Complete the following reading assignment: [1] Jessica Dai, Paula Gradu, Inioluwa Deborah Raji, Benjamin Recht. From Individual Experience to Collective Evidence: A Reporting-Based Framework for Identifying Systemic Harms. ICML 2025 [2] Ben Chugg, Santiago Cortes-Gomez, Bryan Wilder, Aaditya Ramdas. Auditing Fairness by Betting. NeurIPS 2023.
  • Previous Project

Coding Exercises for Teaching Probability Theory
Peter Chi • pbchi@ucsd.edu
TA: TBA
D14 8 seats Probably sometime on MWF afternoon


Within the past two decades, simulation-based inference has established itself as a standard approach for teaching an introductory statistics or data science course (such as DSC 10). While it has been argued that simulation-based pedagogies should likewise be useful in a probability theory course, the implementation therein is not currently well developed, nor is this notion even universally accepted to date. Students in this domain will explore this by developing coding exercises that are designed to teach concepts from a typical undergraduate probability theory course (such as MATH 180A, MATH 183, and MATH 181A). Specifically, the coding exercises that capstone students in this domain will create as part of their projects will task probability students with writing simulation code that illustrates a particular concept or theoretical result in a probability course. Several such coding exercise sets have already been developed by past capstone students, but these could be further polished/refined. Thus, possible deliverables at the end of projects in this domain could be each of the following, or other related/comparable items proposed by capstone students in this domain: (1) a set of coding exercises that each target a specific topic in a probability theory course; (2) instructor lesson notes for each coding exercise that details how it could be implemented in a typical course and its pedagogical rationale (i.e. why we believe it should be effective); (3) solutions to each coding exercise, in both R and Python; (4) Shiny apps (written in either R or Python) for each coding exercise that visually and interactively demonstrate the solution code in action; (5) assessment questions to test students on their resulting understanding of the probability concepts aimed to be taught by each of the coding exercises, and their answers; (6) editing of previously created exercises; (7) an organized website with links to access each of the coding exercises and other public-facing materials.
Read more
  • About: I received my Masters in Genetic Epidemiology from Johns Hopkins University, and PhD in Biostatistics from the University of Washington. My research interests are in statistical phylogenetics, statistics education, and statistics of casino games. Outside of work, I like to hang out with my five year old daughter and practice my Mandarin Chinese.
  • Mentoring Style: I will mentor this domain solo (without assistance from graduate students). I aim to give students the background and confidence to take ownership over their projects, and will likely be fairly hands-on at first: in particular, I will lead discussions on concepts in the field of statistics education, and on probability theory as needed
  • Suggested Prerequisites:
  • Summer Tasks: (1) Review concepts from whichever probability course(s) that you have taken; (2) Explore the Shiny apps here to get an idea of what is possible with Shiny apps for teaching statistics: https://statistics.calpoly.edu/shiny; (3) Read this paper (although its aims were different from what ours will be, it is one of the only scientific research papers to date that addresses the idea of using simulation in a probability theory course, so it will be good to be familiar with what they have done and the issues that they raise): https://www.tandfonline.com/doi/full/10.1080/10691898.2019.1600387
  • Previous Project

Quantization in ML Models
Arya Mazumdar • amazumdar@ucsd.edu
TA: TBA
D16 8 seats Fri 1-2pm


(There is no minimum length of a tweet - so this question is ill-informed). Quantization and Compression of Models that achieves the Information Theoretic Limits. Evaluate performance of these models. Quantization is the process of mapping continuous or high-resolution signals into a finite set of discrete levels. It is fundamental to digital signal processing, data compression, and modern machine learning systems. Despite its apparent simplicity, quantization sits at the intersection of probability theory, optimization, and information theory. This project aims to study quantization through an information-theoretic lens. The central question is: how much information is lost when continuous data is discretized, and how can we design quantizers that minimize this loss under constraints such as bitrate or storage? The project will explore both classical results and modern perspectives, including connections to rate–distortion theory and emerging applications in learning systems.
Read more
  • About: Arya Mazumdar is the HDSI Endowed Chair Professor in AI at University of California San Diego. Arya received his Ph.D. degree from the University of Maryland, College Park, in 2011. Subsequently, he was a postdoctoral researcher with the Massachusetts Institute of Technology. In the past, he has been a professor in University of Minnesota and University of Massachusetts Amherst, and a research scientist in Amazon. His research areas cover algorithmic and statistical aspects of machine learning, error correcting codes, optimization, and signal processing. Awards received by Arya include the 2011 ECE Distinguished Dissertation Award and 2025 ECE Distinguished Alumni Award (both from Univ. of Maryland), EURASIP Best Paper Award, IEEE ISIT Jack Keil Wolf paper award, and an NSF CAREER Award. He was a Distinguished Lecturer of the Information Theory Society for 2023-24, and currently serves as an associate editor for IEEE Transactions on Information Theory, and an area editor for Now Publishers Foundation and Trends in Communication and Information Theory series, among others.
  • Mentoring Style: Work with PhD students and Postdocs if available
  • Suggested Prerequisites:
  • Summer Tasks: Read up quantization papers
  • Previous Project

Applied Data Science

(back to the outline)

ALERTCalifornia: Extreme Events Detection
Nathan Hui, Falko Kuester, Neal Driscoll • nthui@ucsd.edu, fkuester@ucsd.edu, ndriscoll@ucsd.edu
TA: TBA
D01 8 seats M/Tu/We/Th/Fr 10a-11a


The ALERTCalifornia research program continues UC San Diego’s more than 20-year legacy of collecting high-quality data through a network of natural hazard monitoring and detection cameras across the state. This growing network includes over 1,200 camera sensors that provide real-time imagery. They are located in wild spaces, on towers, and other high points across the entire state of California, and are used to watch for and monitor extreme events including wildfire and weather. The program’s historical archive of camera data contains over 40 billion timestamped and localized frames. These camera data have facilitated CALFIRE’s ability to rapidly respond to emerging wildfires as well as maintain situational awareness during ongoing wildfires and other natural disasters. We would like to investigate where machine learning techniques can assist with assessing camera network health, data integrity, and environmental signals. Potential projects include camera site uptime detection, cloud detection, marine layer height detection, Visual Flight Rules altitude estimation, horizon detection, camera positioning calibration, digital twin development, and semantic search of archived camera data.
Read more
  • About: Nathan Hui is currently a research engineer at UC San Diego at the Qualcomm Institute. His area of focus is multi-domain robotics, 3D imaging, and distributed sensor networks. Previous projects include tracking transmittered wildlife using drones, measuring physical oceanographic data using intelligent surfboard fins, and measuring fish length using low-cost lasers, dive cameras, and machine learning. Prof. Kuester received an MS degree in Mechanical Engineering in 1994 and MS degree in Computer Science and Engineering in 1995 from the University of Michigan, Ann Arbor. In 2001 he received a Ph.D. from the University of California, Davis and currently is the Calit2 Professor for Visualization and Virtual Reality at the University of California, San Diego. Professor Kuester holds appointments as Professor in the Departments of Structural Engineering and Computer Science and Engineering at the Jacobs School of Engineering (JSoE) and serves as the director of the Cultural Heritage Engineering Initiative (CHEI), the Center of Interdisciplinary Science for Art, Architecture and Archaeology (CISA3), the Calit2 Center of Graphics, Visualization and Virtual Reality (GRAVITY) and the DroneLab. Neal Driscoll is a professor of geology and geophysics in the Geosciences Research Division at Scripps Institution of Oceanography at UC San Diego. Driscoll researches tectonic deformation and the evolution of landscapes and seascapes. His work primarily focuses on the sediment record to understand the processes that shaped the earth. As part of this research, Driscoll spends time at sea acquiring images of the seafloor and subsurface layers to understand the processes that shape Earth. Driscoll is also co-director of UC San Diego’s Center for Public Preparedness (CP2) and the ALERTCalifornia public safety program. ALERTCalifornia provides critical infrastructure for mitigating wildfire and natural disaster risk to life, property and ecosystems. The advanced network of more than 1200 cameras across California helps first responders monitor natural disasters such as wildfires, floods, and landslides. ALERTCalifornia is a vital resource that provides an array of technological tools, infrastructure and research that supports government agencies, utilities and the public in their response to ever-increasing natural disaster risk. ALERTCalifornia also gathers vital data to inform the greater understanding of natural disaster causes, active event behavior and post-event impacts to air quality, water quality, ecosystems, and human health.
  • Mentoring Style: We can facilitate mentorship in our facilities (Atkinson Hall). This will occur as part of our research group (regular meetings), with additional oversight under ALERTCalifornia (milestone updates).
  • Suggested Prerequisites:
  • Summer Tasks: Students should be able to utilize Nautilus NRP, be familiar with active learning techniques, semi-supervised or unsupervised learning, and utilizing web APIs. We expect most work to be done in Python. Please also be familiar with Docker, Poetry, Kubernetes, and Tornado.
  • Previous Project

analysis of lab mouse video data to automate detection of drug effects
ben smarr, manny ruidiaz from TLR Ventures • manny.ruidiaz@murine.org, bsmarr@ucsd.edu
TA: TBA
D02 4 seats Wed PM


In this work we use mutlimodal time series data extracted from videos of laboratory mice experiencing different pharmacological treatments. The time series reveal patterns like daily and estrual rhythms, and how those rhythms may be perturbed by the treatments. Current assessments of the impact of pharmacological substances is done by invasive behavioral testing which is time consuming, shows substantial inter-scorer error (strongly depends on who ran which tests), and is expensive, time consuming, and bothersome to the animals. The intention is to generate automated pipelines that detect longer and more subtle effects, while being more efficient and less disruptive to the animals. This in turn would support work by drug development teams to more efficiently detect potential signs of responders and non-responders, as well as to identify possible side effects and long-term disruptions not visible through single time-point assessments.
Read more
  • About: Prof Smarr works on time series analyses in biological systems. His work focuses on algorithmic equity, and especially how people and animals show differences in temporal regulation or dynamics that color analyses on data from those subjects. These data-driven identities often differ from classical demographics (e.g. not all women show stable menstrual cycles, so "cycle type" is not the same as "sex") and allow for better accounting of variance in biological system data, enabling greater precision in health predictive algorithms.
  • Mentoring Style: Students will work on weekly tasks, developing comfort with the data and analytic techniques, as well as gaining experience designing and scoping appropriate experiments. I seek to facilitate student-driven work, and not to micromanage. I expect students to be interested in learning how to own a project and impress a private sector stake holder. This is excellent experience for job hunts and consulting work, and my goal is to have students leave with the ability to confidently and professionally describe and defend their work. I prefer in person discussions but we will mix in zoom updates and meetings with industry partners as needed.
  • Suggested Prerequisites:
  • Summer Tasks: Read and recreate figures from the following 3 papers to develop some familiarity with both the behavioral phenomena being analyzed, and the approaches to this analysis. Data and code (not probably the best to use, but useful for reference) are attached to the papers. 1. https://pubmed.ncbi.nlm.nih.gov/27467519/ - analysis of mouse pregnancy 2. https://pubmed.ncbi.nlm.nih.gov/28203366/ - analysis of sex differences by time 3. https://pubmed.ncbi.nlm.nih.gov/35870975/ - expanded analysis of sex differences by time
  • Previous Project

Polygenic Precision Genomics
Tiffany Amariuta • tamariutabartell@ucsd.edu
TA: TBA
D03 8 seats Tuesdays 11 am


Genetic risk prediction is a powerful tool for disease prediction and secondary prevention and has imminent clinical relevance. Students will use population genetics and genomics data to assess individual risk for disease outcomes and transcriptomic measurements. Students will learn to work with genotype data from the 1000 Genomes Consortium and genetic association data from genome-wide association studies (GWAS) and transcriptome-wide association studies (TWAS). The goal of our projects will be to understand the genetic variants regulating gene expression and the genes regulating complex diseases, in an effort to nominate drug targets.
Read more
  • About: Before starting her lab in San Diego, Tiffany earned a B.S. in Biological Engineering at MIT and PhD from the Bioinformatics and Integrative Genomics program at Harvard Medical School, where she studied the genetic susceptibility of autoimmune diseases and other polygenic diseases. She pursued post-doctoral research studying tissue-mediated genetic effects at the Harvard School of Public Health. Now, Tiffany is an Assistant Professor in the Halıcıoğlu Data Science Institute and the Department of Medicine at the University of California San Diego. In her free time, Tiffany enjoys hiking, tennis, beach volleyball, and spending time with her dog, Dax.
  • Mentoring Style: I will be the sole mentor, so students will work directly with me. During the first quarter, each student works individually but collaboratively (everyone submits their own assignments). During the second quarter, we will break into two teams to work on a larger, more innovative project.
  • Suggested Prerequisites:
  • Summer Tasks: 1. Become familiar with coding in R (data visualization, reading in datasets, perform matrix math and other linear algebra operations) 2. Become familiar with coding in unix for DSMLP cluster. 3. GTEx gene expression dataset paper: https://www.science.org/doi/10.1126/science.aaz1776 4. 1000 Genomes/HapMap gene expression dataset paper: https://pubmed.ncbi.nlm.nih.gov/24037378/
  • Previous Project

Analysis of high-dimensional neural data
Mikio Aoi • maoi@ucsd.edu
TA: TBA
D06 8 seats Tuesdays 10-10:50


Neural data is increasingly high dimensional. For example, neuroscientists are routinely recording from thousands of neurons simultaneously and that yield is doubled every 6 years. Similarly, there are now full anatomical connectomes of multiple animals constituting 139,255 neurons and over 50 million synaptic connections. In order to make sense of these data we need principled ways of analyzing them. This capstone will introduce two approaches-- latent variable models, and distance-based analysis which are now used in the latest research studies. Students will learn the basic mathematics of these approaches and some of the latest applications as preparation for developing their own analyses of publicly available datasets.
Read more
  • About: Mikio Aoi is a computational neuroscientist and an Assistant Professor at UCSD, jointly appointed between the Department of Neurobiology and the Halıcıoğlu Data Science Institute. He studies how populations of neurons coordinate their activity to determine the representational and algorithmic functions of neural computations. In particular, his interests are in understanding how the specifics of neural computations impact behavior and in developing principled approaches to data analysis for this purpose. Before pursuing neuroscience he earned a PhD in Mathematical Biology from North Carolina State University studying the dynamics of cerebrovascular function in stroke and diabetes. He served as a postdoc in the Department of Mathematics at Boston University developing methods for studying rhythmic synchrony in electrophysiological measurements and at Princeton University developing scalable methods for analyzing neuronal population dynamics.
  • Mentoring Style: I'll be the only person giving instructions. I will be very hands-on in the beginning, providing background and orientation but students will need to choose their own projects and be mostly independent by winter quarter.
  • Suggested Prerequisites:
  • Summer Tasks: Please follow this link to a document with reading assignments: https://docs.google.com/document/d/1AvxXZ4hkYnj1nY4_fO1s7cP-GI5TZjIGIcrq4bAYbnE/edit?usp=sharing
  • Previous Project

Tritons Roster Optimization Project
Rayan Saab (main), Bradley Voytek • rsaab@ucsd.edu
TA: TBA
D11 5 seats As of now: Monday 2 pm


Roster construction is an inherently uncertain and complex process. For some NCAA sports, including basketball, players can opt to enter their name into the NCAA transfer portal. Once the portal opens, coaches can begin contacting and recruiting transferring players. This is an incredibly fast moving process, where news of athletes leaving and joining teams develops quickly. It is similar to free agency in professional sports, but operates on a much shorter timeline. Teams rarely know in advance which of their current players will remain and which will leave. This uncertainty makes it difficult to design recruitment strategies that maintain team performance while staying within budget constraints. From a mathematical perspective, the problem involves decision making under uncertainty, constrained optimization, and the modeling of many possible roster configurations. This project proposes a data science framework that systematically simulates possible roster departure scenarios and identifies effective recruitment strategies for each case. The goal is to develop a decision support tool that helps teams anticipate roster gaps, quantify tradeoffs, and proactively identify candidate player profiles that best fill those gaps.
Read more
  • About: Rayan Saab is a Professor in the Mathematics Department and at the Halicioglu Data Science Institute. He works on developing computational methods and theory for solving problems related to collecting, processing, and analyzing data. He came to this work first through an undergrad degree in electrical engineering and finding himself always interested in both making things work and understanding why they do.
  • Mentoring Style: I am relatively hands-on in the sense that I make myself available for problem-solving and discussions. That said, students have to be self-motivated, and motivated to do the readings and the work.
  • Suggested Prerequisites:
  • Summer Tasks: TBD
  • Previous Project

Learning to map sounds into body actions
Victor Minces • vminces@ucsd.edu
TA: TBA
D13 6 seats Mondays or Tuesdays will be best. Any time after 10 am.


We will create an interface for users to train a model to map sounds into body gestures, then the users will create sounds with their body gestures. With this tool we can create an application for the speech impaired to be able to talk, or we can create a tool for dancers and artists. The body gestures can be face position, hand position, or body position. For example, a person can move their lips in unison with a voice recording, the model will learn to map lip position into sound, then when the person moves their lips, the model will generate the sound. For another example, a dancer can dance as a musician plays, this trains the model, then when the dancer moves you can hear new music.
Read more
  • About: I have a background in fine arts, physics, and computational neuroscience. I have now combined all these passions. I create web applications for people to play, create, and learn about sound, they are used a lot all over the country. I am also a sound artist, I have made public interactive sound sculptures in public places such as Balboa Park. I love everything sound.
  • Mentoring Style: I love this project so I want to move it forward. We will be developing this project from a scratch. We will work collaboratively. We will build the tool in small and clear steps. We will meet weekly, and I'll be happy to meet at other times. Using Claude Code or Codex is not only fine but encouraged.
  • Suggested Prerequisites:
  • Summer Tasks: -Watch tons of videos about sound generation. -Understand frequency decomposition (Fourier transforms). -Watch this video: https://www.youtube.com/watch?v=IkuzquGec4Y. -If you can, Download a pre-trained model from the RAVE GitHub Run the provided inference examples to hear what it does. -Let me know if you have trouble. I'm open to meet during the summer.
  • Previous Project

The value of prediction for resource allocation
Emily Aiken • eaiken@ucsd.edu
TA: TBA
D17 6 seats Tuesday, Wednesday, or Thursday afternoon


Students in this capstone group will work on applying the "relative value of prediction" framework (https://arxiv.org/abs/2602.08786) to resource allocation problems in public policy. Students will work with the PI to obtain a publicly available dataset relative to public policy and train machine learning algorithms to predict need for interventions at the individual or household level. They will then calculate cost-efficiency trade-offs in the value of prediction relative to other policy makers available in the application domain they have chosen. This work will draw on methods from both data science and economics, as well as domain expertise in different areas of public policy (e.g. healthcare, environment, employment, education).
Read more
  • About: I am an assistant professor at UC San Diego, jointly appointed in the School of Global Policy and Strategy and the Halıcıoğlu Data Science Institute. My research interests are in data science and development economics, with a focus on analyzing large digital traces to inform the design and targeting of social protection and humanitarian aid programs. I was previously a postdoctoral scholar at Carnegie Mellon University Africa (2024-2025). I received my PhD from the UC Berkeley School of Information, where I was advised by Joshua Blumenstock. I also hold an MS (UC Berkeley) and BA (Harvard) in computer science.
  • Mentoring Style: My capstone students will work directly with me this year (not with my PhD students or postdocs). I will be fairly hands-on, and will expect weekly progress reports prior to the capstone meeting.
  • Suggested Prerequisites:
  • Summer Tasks: (1) Read the following papers: https://proceedings.mlr.press/v267/fischer-abaigar25a.html, https://arxiv.org/abs/2602.08786 (2) Provide links to three publicly available datasets that could be used for real world prediction tasks in domains you are interested in
  • Previous Project

Learning and Using Causal Relationships from Multi-modality Data
Biwei Huang •
TA: TBA
D21 8 seats Monday 5 pm - 6pm


Causal discovery and causal AI aim to move beyond pattern recognition by discovering the underlying cause-and-effect relationships that generate observed data. While traditional AI often learns correlations, causal AI seeks to understand why things happen, what would change under an intervention, and how systems behave in new or shifting environments. Causal discovery provides methods for identifying causal structures from observational data. Together, causal discovery and causal AI enable more reliable reasoning, robust prediction, decision-making, scientific discovery, and trustworthy AI systems that can answer “what if” questions and support actions rather than merely describe associations.
Read more
  • About: I am Biwei Huang, an assistant professor at HDSI. My research interests are mainly in causality, including causal discovery, causal AI, and causal foundation models.
  • Mentoring Style: capstone students will work on the projects as a separate entity, mentored by myself
  • Suggested Prerequisites:
  • Summer Tasks: read this paper: https://arxiv.org/pdf/2504.13263
  • Previous Project

Physical Intelligence for Intelligent Manufacturing
Tauhidur Rahman • trahman@ucsd.edu
TA: TBA
D22 8 seats one hour per week


The global manufacturing sector is undergoing a profound transformation driven by the convergence of artificial intelligence, robotics, and advanced sensing technologies. Despite decades of automation, modern factories still face three fundamental and largely unsolved problems: sensing systems that are too slow to perceive physical defect events that occur at sub-millisecond timescales, robots that must be entirely reprogrammed whenever a new task is introduced because they cannot accumulate knowledge over time, and quality control pipelines that remain disconnected from the machines that produce defects in the first place. This domain broadly concerns the question of how autonomous physical agents can be built to sense efficiently, learn continuously, and act precisely in real manufacturing environments. Students will explore ideas across reinforcement learning for physical systems, neuromorphic and event-driven sensing, continual and lifelong machine learning, tactile and multimodal sensor fusion, and closed-loop process control. The domain draws on biological principles, particularly the architecture of the mammalian nervous system, as inspiration for building robotic systems that mirror the way living creatures allocate attention, consolidate memory, and refine motor skill over a lifetime of experience. This is not a domain with settled answers; it sits at the frontier of what AI and robotics research can currently achieve, and students should expect to engage seriously with open problems rather than apply existing off-the-shelf solutions. Relevant entry points into the literature include Gallego et al. (2022) "Event-Based Vision: A Survey" in IEEE Transactions on Pattern Analysis and Machine Intelligence, Wołczyk et al. (2021) on the ContinualWorld benchmark (NeurIPS 2021, https://continualworld.github.io), and LeCun's "A Path Towards Autonomous Machine Intelligence" (2022, https://openreview.net/forum?id=BZ5a1r-kVsf).
Read more
  • About: https://www.tauhidurrahman.com
  • Mentoring Style: https://mosaic-laboratory.github.io
  • Suggested Prerequisites:
  • Summer Tasks: Read relevant papers
  • Previous Project

AI for Good : Misinformation Disinformation Detection, Ranking and Mitgation
Ali Arsanjani • arsanjani@google.com
TA: TBA
D24 16 seats true Industry Partner


currently the world of media and constant is experiencing a constant barage of disinformation, and fake news, in this project we will teach generative and agentic ai and explore how to use them to help categorize, detect, rank and mitigate fake news ?
Read more
  • About: Dr. Ali Arsanjani is the Director of Applied AI Engineering at Google Cloud, and Head of the Global AI Center of Excellence, leads GenAI strategic co-innovation, thought leadership & alliances. Previously he led all GenAI initiatives in the Google Partner Ecosystem, developing strategic co-innovation assets & partnerships in the fields of Generative AI, Data/Analytics & Predictive AI/ML. His team specializes in co-innovation with ISV and GSI partners as they run, integrate and build on GCP across the ML Lifecycle. Ali also works closely with product management to shape the direction of Google's AI and analytics offerings from a cloud perspective. Ali is an Adjunct Prof at San Jose State University & the University of California, San Diego, and advises students in the Masters in Data program and the Data Science Institute, respectively. Prev, Ali was the Head of Machine Learning & Worldwide TechLeader for AWS AI / ML specialist solution architecture. As Chief Principal AI/ML, he developed tech strategy, led the Tech Sector for Saas, ISVs, and the internet (AWS's largest customers). He led research, product and eng teams responsible for prototyping cutting-edge AI devices. Ali was founder of AI startup Deep Context, which provides clients with actionable data on contextually relevant customer engagement. As VP of AI / ML at 8x8, a cloud communications company, Ali was responsible for defining the strategy, research direction, product management, and implementation of AI and ML products in the telecommunications sector, with a focus on unified communications as a service and contact center as a service products. From 1998 to 2018, Ali worked at IBM as CTO of Analytics & ML and a Distinguished Engineer, leading a team of 6,000 practitioners worldwide in the development of bespoke ML and analytics solutions. He was responsible for leading in SOA, BPM, RPA, analytics, AI/ML systems and solutions. Ali was recognized as the "father of service-oriented architecture" within IBM and held the title of Master Inventor, leading the Global Invention Disclosure Team and the Core Management Group for the IBM Academy of Technology. Ali's expertise includes NLP, LLMs, deep learning ensemble models, anomaly and outlier detection and training, customer segmentation and analysis for churn and upselling, and voice, video, and text analysis for conversational virtual assistant implementations. He has also served as a chair for standard bodies such as The Open Group, co-leading the development of standards for SOA Ref Arch, maturity model, and cloud computing arch.
  • Mentoring Style: This will be a classroom style augmented with standups
  • Suggested Prerequisites:
  • Summer Tasks: alternusvera.com is a pre read. I will also send papers
  • Previous Project

Deep Learning for Climate Model Emulation
Duncan Watson-Parris • dwatsonparris@ucsd.edu
TA: TBA
D26 8 seats Most likely a M/W/F morning


The choices humanity makes in the next few decades will determine how much warmer the Earth will be by the end of the century, with implications for billions of lives and trillions of dollars in GDP. Many different emission pathways are compatible with the Paris climate agreement, and many more miss that target. Full-complexity Earth System Models can only simulate a small subset of these scenarios, so fast, accurate emulators are essential for exploring the space of possibilities and quantifying the associated risks. Our lab developed ClimateBench (github.com/duncanwp/ClimateBench) and more recently ClimateBench2 as a benchmark for this problem that unlock the possibility of more accurate and skillful climate models. This project will develop direct climate emulators in JAX that targets these datasets and integrates with our differentiable JAX Earth Model (JEM). Students will explore architectures (neural operators, transformers, pattern-scaling hybrids), physical constraints (energy conservation, spatial coherence), and the use of jax.grad for gradient-based scenario analysis and calibration. The goal is a new generation of emulators that are fast, reliable, differentiable, and suitable for integration into downstream decision-support.
Read more
  • About: Duncan Watson-Parris is an atmospheric physicist working at the interface of climate research and machine learning. The Climate Analytics Lab (CAL) at Scripps / HDSI focuses on aerosol-cloud interactions, their representation in global climate models, and the development of differentiable and hybrid ML models for climate science. CAL leads a variety of open-source ML tools and convenes the Machine Learning for Climate Science EGU session and the "AI and Climate Science" discovery series as part of the UN AI for Good program.
  • Mentoring Style: This work is central to CAL, so I will meet with the students directly and the project can be integrated into the broader lab research program to the extent students wish. Students are welcome to join our lab meetings at Scripps Institution of Oceanography.
  • Suggested Prerequisites:
  • Summer Tasks: - Skim the IPCC AR6 Synthesis Report, especially the SPM figures: https://www.ipcc.ch/report/ar6/syr/downloads/report/IPCC_AR6_SYR_SPM.pdf - Read the ClimateBench paper (Watson-Parris et al. 2022): https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2021MS002954 - Work through the JAX "Getting Started" tutorials and the Flax NNX quickstart: https://docs.jax.dev and https://flax.readthedocs.io - Get comfortable with xarray for gridded climate data: https://docs.xarray.dev/en/stable/
  • Previous Project

From First Search to Trust at Scale: Building Data-Driven Marketing Systems and AI Agents for the Fertility Patient Journey and Online Reputation Management
Gabriel Andres Garzo and Madison Hughes • gabrielandres.garzo@rpsandiego.com, madison.hughes@rpsandiego.com
TA: TBA
D27 8 seats Tuesdays at 3:30pm PST Industry Partner


Our domain sits at the intersection of healthcare, marketing, and patient access, specifically within the fertility space where individuals often face complex medical, financial, and emotional barriers to care. Infertility affects millions of people, yet awareness, education, and access remain highly fragmented due to factors such as insurance coverage, socioeconomic status, cultural stigma, and rapidly evolving medical technologies. A central challenge in this space is understanding how patients move through their journey, from initial awareness and research to selecting a provider and beginning treatment, and identifying where friction or drop-off occurs. This includes analyzing digital behavior across search, social media, and AI-driven platforms, as well as referral patterns from physicians and insurance networks, and the role of trust signals like reviews, educational content, and community engagement. An emerging focus within this domain is the development of intelligent systems or agents that can oversee and manage a clinic’s online reputation at scale, including monitoring reviews, identifying sentiment trends, generating appropriate responses, and surfacing actionable insights to improve patient experience and trust. At the same time, organizations must navigate strict regulatory and ethical considerations while supporting patients through deeply personal experiences and continuing to drive growth and access to care. Students can explore topics such as patient decision-making in healthcare, access and policy including SB 729 California fertility coverage law, digital health marketing trends, online reputation management, and disparities in reproductive care. Helpful resources include RPSD's website, RESOLVE: The National Infertility Association for advocacy and patient perspectives, American Society for Reproductive Medicine for clinical and research insights, and broader literature on healthcare consumer behavior and digital transformation in medicine.
Read more
  • About: We represent both the leadership and marketing strategy behind Reproductive Partners Fertility Center in San Diego. Gabriel Andres Garzo is the Executive Director with a focus on expanding access to fertility care, navigating evolving insurance landscapes, and driving operational growth. Alongside this, Maddie Hughes serves as Marketing Manager, having built the center’s marketing function from the ground up, leading initiatives across digital strategy, patient acquisition, partnerships, and community engagement. Together, our shared interest lies in improving access to fertility care through a combination of data-driven decision making, patient-centered storytelling, and innovative outreach. We are particularly focused on how marketing, operations, and policy intersect to shape the patient journey from first touchpoint through treatment.
  • Mentoring Style: This capstone will be run as a focused, industry-style working group rather than part of a larger research lab. Students will work directly with Gabriel and Maddie and will be treated as a small, collaborative team solving real-world marketing and analytics problems in the healthcare space. We will be very hands-on, especially in the early stages, helping to frame the problem space, guide research direction, and ensure students are thinking about both the technical and business implications of their work. As the quarter progresses, the expectation is that students become increasingly independent, taking ownership of their projects, driving analysis, and forming recommendations with less day-to-day direction. Students can expect a mix of structured guidance and autonomy. We will provide clear expectations, timelines, and regular feedback, along with examples of how similar problems are approached in practice. At the same time, there will be an emphasis on critical thinking, problem solving, and initiative. This is not a step-by-step project. Students will be expected to navigate ambiguity, ask thoughtful questions, and iterate their ideas based on feedback. There will be a strong focus on translating data into actionable insights. In addition to technical work, students will be expected to communicate their findings clearly, present to the group, and think through how their recommendations would be implemented in a real organization. We will be actively involved in weekly sessions, project reviews, and office hours, and will aim to create an environment that is supportive, fast-paced, and reflective of how data and marketing teams operate in a professional setting.
  • Suggested Prerequisites:
  • Summer Tasks: Students entering this domain will benefit from building a strong foundation in healthcare marketing, patient behavior, and digital analytics prior to the start of the program. Over the summer, we recommend exploring key concepts such as marketing funnels, attribution modeling, A/B testing, and SEO/paid media strategy, as well as how these apply in regulated industries like healthcare. To better understand the fertility space and patient perspective, students should review resources from RESOLVE: The National Infertility Association and clinical overviews from American Society for Reproductive Medicine. For marketing and growth strategy, foundational readings such as “Traction” by Gabriel Weinberg and Justin Mares and “Hacking Growth” by Sean Ellis and Morgan Brown are helpful, along with articles on healthcare consumer behavior and digital transformation in medicine. From a technical perspective, students should be comfortable working with Python, SQL, and data visualization tools such as MS Power BI, and should review concepts in exploratory data analysis and hypothesis testing. Familiarity with tools like Google Analytics, CRM platforms (Salesforce Sales Cloud and Salesforce Marketing Cloud), or marketing automation systems is a plus but not required. For Quarter 1, students may begin thinking about how to replicate and analyze a marketing funnel or attribution model using sample or publicly available datasets, with a focus on identifying drop-off points and opportunities for optimization. More broadly, students should come in prepared to think critically about how data can be used to improve real-world decision making in a sensitive, patient-centered context.
  • Previous Project

NLP Credit Score Development
Kyle Nero, Daniel Mathew, Berk Ustun • kyle.nero@prismdata.com, daniel.mathew@prismdata.com, berk@ucsd.edu
TA: TBA
D28 12 seats Thursday 1pm or 2pm


One of the most widely used and little understood parts of the Financial Services industry is the credit score. In this course, students will work with transactional bank data to build statistical models for the purpose of assessing creditworthiness in the financial services industry. The course will take students through the life of a model development project, from data exploration, through model training and evaluation. Students will have the opportunity to work with both structured and unstructured data as they learn about the process and attributes that go into credit scores. Additionally, students will learn about the importance of model explainability and fairness.
Read more
  • About: Kyle: Kyle Nero graduated from UCSD HDSI in 2023, majoring in data science and minoring in business. During his senior year, he engaged with industry partner Prism Data through the HDSI Senior Capstone Project. He went on to intern with Prism Data following his graduation and joined the team full time as a Data Scientist in September 2023. Daniel: Daniel Mathew graduated from UCSD HDSI in 2025. He also engaged with Prism Data through the Senior Capstone Project and went on to join the Prism team as a Data Science Intern and converted to a full time Data Scientist in September 2025.
  • Mentoring Style: Our group will consist of group projects completed in groups of 3-4. The goal of the course is to eventually build a credit score but we will start by building a transaction categorization model using NLP techniques. Each week we will talk about techniques that can be applied to the next step in the project. We will begin by reviewing homework from the previous week and discussing ideas. Then introduce the next step and talk about what can be done to solve the next step in the problem. The goal is to introduce students to the model development process in most financial services companies.
  • Suggested Prerequisites:
  • Summer Tasks: https://www.capitalone.com/learn-grow/money-management/when-did-credit-scores-start/, https://www.capitalone.com/learn-grow/money-management/fair-credit-reporting-act/, https://www.capitalone.com/learn-grow/money-management/equal-credit-opportunity-act/, https://www.nerdwallet.com/article/finance/credit-score-ranges-and-how-to-improve and https://developers.google.com/machine-learning/crash-course/classification/roc-and-auc#:~:text=An%20ROC%20curve%20(receiver%20operating,False%20Positive%20Rate
  • Previous Project

(1) European Wildfire and Property Intelligence Modeling with Cotality (2) Property Intelligence: Enhancing Geospatial Data Quality for Risk Assessment
Ilyes Meftah, Justin Murphy, Mark Dixon, Lawrence Vulis • imeftah@cotality.com, jmurphy@cotality.com, markdixon@cotality.com, lvulis@cotality.com
TA: TBA
D31 8 seats Wildfire: As 2 of the mentors are UK-based (PST-8 hours) we suggest Mondays at 8am PST. For property domain (2) Thursday or Friday mornings (9 AM) is good. Industry Partner


(1) This capstone will offer students the opportunity to apply advanced data science methods to model European wildfire risk in collaboration with Cotality. Over the last decade, both the number and intensity of major wildfires impacting Europe has significantly increased, impacting both ecosystems and communities. The aim of this capstone is to produce spatial wildfire probability maps by combining historical observations with key meteorological parameters obtained using state-of-the-art numerical weather model reanalyses. Such maps can be used to quantify the socioeconomic risks posed by wildfire. Challenges include dealing with large data sets; identifying and correcting trends in historical data (including possible climate change signals); determining efficient methods of detecting statistical relationships between observations and independent parameters; developing spatial analysis methods. Students will obtain a solid grounding in domain knowledge covering the science of wildfires and associated meteorology, as well as gaining an understanding of the data science processes involved in building risk products and catastrophe models produced by companies such as Cotality. (2) This domain focuses on solving critical data quality challenges in property analytics that underpin risk assessment models across multiple industries. Students will work with Cotality, a leading provider of property information and analytics, to enhance nationwide property databases through advanced data science techniques. A key part of Cotality's product is providing standardized property characteristics like age of building, land use, square footage, for every building in the United States. This is used downstream for various purposes including catastrophe risk modelling, tax, and other data products. This domain will focus on the detection and correction of anomalous building characteristics that are inherited from various data sources. This will involve including geospatial algorithms for data alignment between various spatial datasets, multiscale outlier detection, and characteristic standardization or dimensionality reduction. In the first quarter students will become familiar with the property data ecosystem and geospatial data science. In the second quarter they will focus on geospatial algorithms for data alignment between various spatial datasets, multiscale outlier detection, and standardization or dimensionality reduction of disparate characteristics. This work directly impacts catastrophe risk modelling, climate risk assessment, mortgage analytics, and urban planning. The project provides hands-on experience with large-scale geospatial datasets, feature engineering, classification algorithms, and data quality methodologies. Students will develop transferable skills in data cleaning, normalization, feature engineering, and model deployment while producing solutions that offer "cross-model impact affecting all catastrophe models and climate-related projects" within the industry. The outcomes will enhance visualization capabilities and improve risk assessment accuracy for multiple stakeholders, including homeowners, financial institutions, and government agencies.
Read more
  • About: Ilyes Meftah has been a data scientist and catastrophe modeler with Cotality for 13 years. With a strong background in mathematics and quantitative finance (holding multiple master's degrees from Paris, France universities), Ilyes has developed risk assessment models for wildfires, hurricanes, and earthquakes throughout his career. Recently, he has been focusing his efforts on quantifying wildfire mitigation measures to help communities located in high-risk areas. He is passionate about solving complex problems and sharing knowledge with others. When not working on catastrophe models, he enjoys hiking around the world with his family. Justin Murphy has been a hazard scientist and catastrophe model developer with Cotality for over 20 years, he is based in the UK. Combining a background in mathematical modelling and a keen interest in natural catastrophes, solving problems and delivering complex projects, he has a wealth of domain expertise developing models across the globe. As a senior leader in the team he is focused on enabling his colleagues to solve complex problems across the model development lifecycle. He is a strong communicator and keen advocate for ongoing learning. His educational background is BSc Applied Mathematics and MSc Mathematical Modelling from UK universities. Justin enjoys hiking and cycling with his wife, often whilst travelling in the mountains of Europe sampling great local food. Mark Dixon – based in the UK - has been a hazard scientist and catastrophe model developer for over 16 years and has been with Cotality for the last 6 years. He has developed risk models spanning many geographical regions and types of natural hazard including windstorms, typhoons, hurricanes, severe storms, flood and wildfire. Previously he worked as a scientist at the UK Met Office on numerical weather prediction models and data assimilation systems. He has a PhD in Condensed Matter physics. Mark enjoys spending time with his family, reading and playing guitar. Lawrence Vulis is a hazard scientist with 3 years of experience at Cotality, where he works on building physical and AI-based models of natural hazard risk to properties. He works on building the property portfolio and modelling flood related perils. This is his second year serving as a mentor at UCSD. Prior to Cotality he worked on satellite imaging projects including classification of river deltas, tracking arctic lake dynamics, tracking Southern California beach dynamics, and permafrost mapping. His educational background is in Civil and Environmental Engineering, with a B.E. from The City College of New York and a Ph.D. from UC Irvine, with an extended internship and brief stint at Los Alamos National Lab. Outside of work he enjoys spending time with his wife and dog on beaches and trails.
  • Mentoring Style: Our mentoring approach combines structure with creativity in a collaborative environment. Weekly sessions will balance technical guidance with hands-on problem-solving. Students will have opportunities to interact with multiple catastrophe modeling experts at Cotality, gaining exposure to different perspectives and specialized knowledge. We believe learning works best when it's engaging and enjoyable, so we'll incorporate real-world applications and team-based challenges throughout the project. While we'll provide regular guidance and feedback, we value student initiative and will encourage independent exploration of solutions within our project framework. Our goal is to create an experience that's both intellectually stimulating and professionally valuable.
  • Suggested Prerequisites:
  • Summer Tasks: Both domains: Suggested preparation: - Spatial Cluster Analysis (YouTube playlist): https://www.youtube.com/playlist?list=PLzREt6r1Nenk3L0ndufhYuwdrrfZqdsIA - Spatial Data Science General Topics (YouTube playlist): https://www.youtube.com/playlist?list=PLzREt6r1NenmFyTw8v2JZpEE4PZGNi5Ht - Python GIS Textbook (Part II and III): https://pythongis.org/part2/index.html - R users: Get comfortable with terra & sf libraries and spatial point pattern analysis More detailed resources will be shared as needed. All students should get an understanding of GIS data from links above. Domain 2: Become familiar with BigQuery/GoogleSQL. Ideally algorithms will be built in SQL. https://docs.cloud.google.com/bigquery/docs/introduction-sql
  • Previous Project

Deep learning techniques for understanding and mitigating electric energy costs at home
Dr. Haojian Jin, Assistant Professor, UCSD; Dr. Michael Deisher, Distinguished Engineer, Azimuth AI; Sylvia Downing, Chief Technology Officer, Azimuth AI • haojian@ucsd.edu, michael.deisher@azimuth-ai.com; sylvia.j.downing@azimuth-ai.com
TA: TBA
D33 8 seats This should be based on Haojian's schedule. He mentioned Wednesday afternoons.


This capstone focuses on understanding and mitigating electrical energy consumption and cost. Students will apply AI techniques to use deploy cases for the electrical grid and endpoints (homes, buildings). This requires understanding of data gathering techniques and privacy requirements. This capstone will explore techniques in anomaly detection, zero shot learning, and machine learning on edge systems. Understanding of basic electricity concepts is desired but not required.
Read more
  • About: Haojian: I am a human-centered system researcher. My research lies at the intersection of human-computer interaction, privacy and security, and systems, focusing primarily on two conflicting questions: How can we design systems that meaningfully improve people’s lives? How can we prevent those systems from becoming unsafe as they grow more proactive and powerful? Mike and Sylvia: Azimuth AI is a custom semiconductor company focused on providing rich computation for edge applications, with a focus on reducing greenhouse gasses. Mike has over 25 years of experience deploying applied AI solutions for sensing, audio, and speech recognition, and most recently electrical metering and battery management. Sylvia has over 35 years of experience introducing new technologies to the market in integrated circuits, including first flash memories, encryption, display and GPU architecture, and multimodal sensing solutions.
  • Mentoring Style: We will meet weekly with students via Zoom. Mike and Sylvia are not in San Diego, so may meet in person a couple of times during the course of the year. However we are available to answer questions as needed via remote means.
  • Suggested Prerequisites:
  • Summer Tasks: It may be helpful to review the following materials on how the electrical grid works: https://understand-energy.stanford.edu/energy-currencies/electricity-grid . It is also useful to have experience with PyTorch.
  • Previous Project

Applying precision medicine to the world's most important patient - the ocean
Rob Knight •
TA: TBA
D34 8 seats 3-4 pm Wednesdays has worked well in the past


We have made tremendous progress on using multi-omics for precision medicine in humans (and even more so, in mice), but application of these techniques to environmental settings has been much more limited. In the eDNAID project, sponsored by the Minderoo Foundation, we are collecting several cutting-edge data modalities (quantitative long-read sequencing on the PacBio HiFi platform, untargeted metabolomics, and metaproteomics on the Thermo Astral mass spectrometry platform) and using them to develop new classifiers of ocean health, as well as a better mechanistic understanding of processes of biological and chemical change in the ocean generally. The focus of this capstone project is to transfer multi-omics analysis techniques and multimodal learning methods that have been successful in biomedical settings to this completely novel dataset, and to pilot what will be possible in human health when the same technology stack is applied to biomedical projects (which is our near-future goal).
Read more
  • About: I trained in biochemistry and ecology & evolution, but with a strong interest in computing. My early work was on the origins of the genetic code, but once I got my faculty position I realized that rather than studying events in the deep past, it was more exciting to affect the future. So the techniques I had been using to study noncoding RNA sequences turned out to apply to the massive populations of microbes that inhabit our bodies and environments, and play many roles we're just starting to uncover. At UCSD I direct the Center for Microbiome Innovation, am a Professor in 4 departments including HDSI, and develop technologies to address large-scale questions at the intersection of human and environmental health.
  • Mentoring Style: Will be integrated with PhD students and postdocs for hands-on work between the weekly class meetings and office hours with me.
  • Suggested Prerequisites:
  • Summer Tasks: Familiarity with PyTorch, skbio, MIINT and RYPE would be very helpful. No prior microbiome or multi-omics experience is needed. Prior experience with large dataset visualization would be helpful.
  • Previous Project

Searching for Dark Matter in Ultra Long Time Series Data
Aobo Li • aol002@ucsd.edu
TA: TBA
D35 5 seats Tuesday 4pm (subject to change)


Dark matter is one of the biggest mystery in modern physics: It makes up 85% of the matter in our universe but has never been detected on earth as of today. Detecting dark matter is really tough because it hardly ever interacts with anything we can easily see or touch. Scientists use special, super-sensitive equipment, such as quantum detectors, to try and spot these elusive particles and learn more about how the universe works. These detectors run continuously, generating an ultra long time series as they wait for a rare dark matter signal to appear. In this project, we will establish an analysis team dedicated to examining this ultra long time series. The team's primary task will be building an advanced neural network to denoise this ultra-long time series data, aiming to filter out the overwhelming background noise and cleanly isolate potential dark matter signals.
Read more
  • About: I am Aobo Li. I am a faculty at HDSI & the Department of Physics. I earned my B.S. from UW Seattle and my PhD from Boston University, both in the field of physics. My research uses machine learning to squeeze out the maximum amount of information from ultra-sensitive radiation detectors, all in the quest to uncover extremely rare physics events in our universe. Beyond academia, my interests span from following e-sports to exploring national parks and photography.
  • Mentoring Style: This year my master student Akbota Assan will lead the Capstone team to explore the TIDMAD dataset. Akbota has been working on TIDMAD and understand the technical subtlety, I will provide overall guidance while Akbota helps address technical questions.
  • Suggested Prerequisites:
  • Summer Tasks: Student should read the TIDMAD paper which describes the dataset: https://openreview.net/pdf?id=GgKHKZgg0Y They should also learn coding neural network in PyTorch, a good resource is on my website: https://aobol.github.io/AoboLi/#tutorials
  • Previous Project

Data-driven design of nonlinear architected materials
Igor Sfiligoi <isfiligoi@sdsc.edu> (SDSC) and Nicholas Boechler <nboechler@ucsd.edu> (MAE) • isfiligoi@sdsc.edu, nboechler@ucsd.edu
TA: TBA
D37 8 seats Mondays, either 9am or 1pm


Throughout history, discoveries of materials with new properties have enabled significant technological leaps. Inspired by this idea, we strive to uncover new understanding that enables the design of materials with extraordinary properties, with a focus on materials wherein new mechanical properties are achieved by manipulating microstructure, topology, and nonlinearity. The SHAPE Center (https://shape.ucsd.edu) is developing a computationally enabled inverse design framework for the design of architected material systems undergoing high energy impacts, with a strong empirical validation component. As part of this we are interested in developing data-driven ML tools to rapidly and accurately parse the architecture design space, as well as gain insights into underlying physical mechanisms via classification. We are particularly interested in the following aspects: 1) ML algorithms that rapidly predict metamaterial architectures for desired nonlinear responses; 2) The use of ML algorithm as an initial seed for a coupled gradient-based optimizer, with the goal of reducing the computational costs compared to pure gradient-based optimization with random initial seeding. Students will have access to both experimental and simulation data.
Read more
  • About: Nick received his B.S. in Aerospace Engineering from Georgia Tech, and his M.S. in Aerospace Engineering and his Ph.D. in Aeronautics from Caltech. He worked at MIT as a Postdoctoral Associate from Fall 2011 to Spring 2013. He has been a faculty since 2013 and is the recipient of the Army Research Office and Air Force Office of Scientific Research Young Investigator Program awards. His research is broadly focused on the design of materials with extraordinary properties, towards the enabling of new technological possibilities. Outside of academia, he has recently co-founded a company (Euler Materials), has two kids (5 and 7) which keep him busy, and when time allows battles weeds in his organic garden and enjoys reading science fiction and fantasy novels. Igor is a SDSC Lead Scientific Software Developer and Researcher who partners with other domain scientists to improve the tools they use. His recent interests and involvement mostly center around GPU and HPC computing, but has been involved in infrastructure development in the past, too. Over the years, he has been working with scientists across structural engineering, computational biology, fusion research, high energy physics and astrophysics. Outside of academia, he likes biking and reading historical fiction. He has a teenage daugther, and likes his garden to look like a tropical island.
  • Mentoring Style: We plan to take an engaged but student-led approach to mentoring. We'll work closely with the students throughout the project - meeting regularly, providing guidance, and being available for feedback and support. That said, we're looking for students who are excited to work independently as much as possible. Closer integration with the SHAPE center graduate student team would be welcome, but it is optional.
  • Suggested Prerequisites:
  • Summer Tasks: Topology optimization, Gradient-based optimization, Engineered metamaterials, 3D printing
  • Previous Project

Fusion Data Science: Databases and AI Models for Gyrokinetic Transport
Igor Sfiligoi; Jeff Candy; Emily Belli • isfiligoi@sdsc.edu; candy@fusion.gat.com; bellie@fusion.gat.com
TA: TBA
D39 8 seats Monday 2pm Industry Partner


In order to make Fusion a practical, zero-carbon energy source for the future, the design of future reactors will require reliable predictive simulations. This means predicting turbulent particle and energy transport from the plasma core to the edge and out of the confinement region. Existing high-fidelity gyrokinetic simulation applications like CGYRO provide first-principles predictions of this transport, but the resulting datasets are expensive to generate, complex, and difficult to use directly in reactor-design or integrated-modeling workflows. Over the years, many such simulation data points have been created, and this project will apply sound data science practices on it. There are two complementary activities that would benefit the Fusion science community; (a) making gyrokinetic simulation data searchable, interpretable, reproducible, curated, and useful for downstream analysis, and (b) using gyrokinetic simulation data to build fast predictive ML models for quantities of interest in fusion transport. We will initially explore both directions, and settle on the most promising path based on student interests.
Read more
  • About: Jeff Candy is a computational plasma physicist and Director of Theory and Computational Science at General Atomics. He specializes in gyrokinetic theory, plasma turbulence, and high-performance simulation for magnetic fusion energy. He received his Ph.D. in Physics from the University of California, San Diego. His professional work focuses on first-principles modeling of turbulent transport in tokamak plasmas, scientific software development, and predictive simulation for fusion-relevant regimes. Outside his fusion research, he has a strong technical interest in theoretical and applied acoustics. He is the co-developer of Speakerbench, a software tool for loudspeaker analysis and design, reflecting his broader interest in applying computational methods to practical audio systems. Emily Belli is a physicist in the Magnetic Fusion Energy group at General Atomics, where she develops mathematical and computational models to advance predictive capability for fusion energy, including running simulations on some of the world's largest exascale computers. She holds a B.S. in Engineering Physics from Cornell University and a Ph.D. in Astrophysical Sciences from Princeton University. Her research is driven by the goal of making fusion a practical, zero-carbon energy source for the future. Outside of work, she enjoys running and has completed several marathons, and is a devoted San Diego Padres fan. Igor Sfiligoi is a SDSC Lead Scientific Software Developer and Researcher who partners with other domain scientists to improve the tools they use. His recent interests and involvement mostly center around GPU and HPC computing, but has been involved in infrastructure development in the past, too. Over the years, he has been working with scientists across structural engineering, computational biology, fusion research, high energy physics and astrophysics. Outside of academia, he likes biking and reading historical fiction. He has a teenage daughter, and likes his garden to look like a tropical island.
  • Mentoring Style: We plan to take an engaged but student-led approach to mentoring. We will work closely with the students throughout the project - meeting regularly, providing guidance, and being available for feedback and support. That said, we're looking for students who are excited to work independently as much as possible.
  • Suggested Prerequisites:
  • Summer Tasks: None Students should review regression, classification, cross-validation, error metrics, uncertainty estimates, and reproducible data workflows. For domain background, students should read introductory material on magnetic-confinement fusion, tokamak transport, plasma turbulence, and the role of gyrokinetic simulations as expensive physics-based calculations that can be used to train or validate faster reduced models. Additional helpful topics include Gaussian-process regression, dimensional analysis, principal-component analysis, uncertainty quantification, and high-performance scientific computing.
  • Previous Project

–>