Capstone Project
Trust and Evidence in AI
Overview
The capstone asks you to begin with a domain problem and intended use, then determine what evidence would justify a specific conclusion about an AI system. The goal is not simply to build a model. It is to identify what the evidence supports, where the system may fail or shift, and what responsibility would remain before real use.
You may work individually or in a team of 2–4 students. Team submissions must include a clear contribution statement. The capstone represents 50% of the course grade.
Timeline and Deliverables
| Deliverable | Due | Weight within capstone | Main format |
|---|---|---|---|
| Project Proposal | September 18, 2026 | 10% | 300–500-word PDF plus one slide, a ~3-minute presentation, and ~4-minute discussion |
| Milestone Report | October 16, 2026 | 20% | At most 3 pages in NeurIPS format; references may be additional |
| Milestone Poster Presentation | October 19, 2026, in class | 10% | One poster per project/team and individual questions |
| Spotlight Presentation | Slides due December 4; presentations December 7 and 9 | 10% | Short final presentation; details will be announced |
| Final Report and Class-Survey Contribution | December 9, 2026 | 50% | At most 5 report pages plus a separate survey contribution |
The live schedule is the source of truth for due dates and class sessions.
Project Categories
Choose the category that best matches the work:
- Application or domain case
- Reproduction, evaluation, and diagnosis
- Survey or white paper
- Theory, algorithm, or system design
The first three categories are recommended for most teams. Large language models and agents are welcome when they serve a clear project question and evaluation; merely using an agent to assemble a pipeline is not enough.
What Every Project Should Make Clear
- The domain decision, intended use, and current non-AI process or baseline
- A specific question and evolving one-sentence thesis
- The evidence, data, sources, and evaluation design
- At least one realistic shift, difficult condition, or unsupported case
- What the results support and what they do not support
- Who would be responsible for review, monitoring, or re-evaluation in practice
- What additional expertise or evidence would be needed before responsible use
Each project also contributes a standardized case to the class survey, currently titled From AI Output to Justified Use: Trust and Evidence Across Domains.