Final Projects: how to begin thinking about yours
Final projects are still a ways off, and we haven’t yet taught you some of the most central tools that you’ll want to use on your final project, so I want to reassure you that you are not behind if you haven’t formed a final project plan yet!
That said, it’s not too early to start contemplating ideas during your “shower thoughts” (or as Markus said, your “walk in the redwoods”) type parts of your day. This page is intended to give you a sense of what the projects might look like, to help give shape to that contemplation, but I want to emphasize that thinking outside the box of what I describe here is also welcome and encouraged!
Project scope and deliverables
- Open-ended, “BYOP”: The final project is intentionally loosely defined, in that our goal is to support you in applying AI to use cases in your field or areas of interest in your work and life. In other words, “Bring Your Own Problem” (BYOP) and we’ll help you craft a solution.
- Projects should be “Agentic,” but that could mean a lot of things! Just like how I said on Day 1 that the term “AI” can mean different things in different contexts, “Agentic” is also a pretty broad, multi-meaning term. This gives you a lot of flexibility in designing your final project to fit your interests and goals. Some of the tools most often associated with “agentic” work are yet to be introduced in our class, but there are some very viable final project ideas that you could tackle with what we’ve already covered.
- Projects should show engagement with ethical issues raised by your use of AI in that context. You should ask yourself, is this an appropriate use of AI? What misuses or risks can I anticipate and what mitigations can I put in place? In the inevitable gaps in the mitigations, how can we understand the distribution of responsibility for things that might go wrong? How could the existence of your project change the context in which it is used, if usage became widespread — what existing community dynamics would shift? Whose work is being used in the creation of your project, and whose work might be disrupted by its use?
- Projects should be something you’ll be excited to share at the end of quarter showcase event, and possibly in a portfolio on GitHub or something like that. The end of quarter showcase event is a “poster session” format event that will be during our final exam timeslot (there is no actual exam). We’ll have guests from around campus and off campus, and it’s a chance to share with them and with your classmates. We’ll have some refreshments as well!
- Deliverables:
- (1) A pre-showcase meeting with your tutorial TA to review the internals of how you implemented your project. This meeting is intended to be a combination of signing off on what you’ve done, and a place where you can ask any last-minute advice questions you have on finishing touches ahead of the showcase. This meeting will make the showcase more of a “victory lap” celebration than real evaluation.
- (2) For the showcase, a poster that presents both the project and your ethics reflections. See Poster Resources for templates and an example.
- (3) For the showcase, a brief (~3-5 minute) oral presentation of your poster. The format of a poster session involves a certain amount of spontaneity in responding to your listener’s questions, but you will want to have a basic “pitch” version that you’ve practiced delivering, so you will feel confident when listeners approach your poster. If you would like to have a laptop out for live demo of your tool, in addition to the poster, that would be a great addition but is not required.
See the Final Project Rubric for how projects are evaluated.
Example ideas
Examples using what you already know
- Create a repository of Gemini “Gems,” Claude “Skills,” or Claude agents.md files and share them with the world. Some ideas within this format:
- Your own version of something like the collection of role-playing game Gems that I mentioned in lecture. (Ethics reflections: (1) Similar to ethics and IP/Copyright concerns around the world of fan fiction, when is it ok to adapt IP that you didn’t originate? To what extend does your work transform the original and to what extent could it be seen as a competing product? (2) If your RPG experience is super-engaging, could you overshoot into creating conditions for harmful addictive behaviors?)
- A collection of course-specific agents.md files for popular Stanford classes. (Ethics reflections: (1) Are you doing this with the buy-in of the faculty who teach these classes and in conformity with their honor code policies? Are you using any of their IP — such as adding documents authored by the instructor into a RAG — and did you get their permission? (2) What responsibilities do students who use the products retain to not try to circumvent the guardrails, who is responsible if the full answers “accidentally” spill out of the system despite the guardrails?)
- A library of “interview me” Gems tailored to specific campus offices (financial aid, study abroad, creating a new student club, hosting an event) that interview a student and draft a first pass at that office’s application. (Ethics reflections: (1) What guardrails can you put into place to ensure that the tool does not circumvent thought processes that some questions on the application/form may be designed to provoke, such as ensuring that the applicant has considered certain policy/safety/etc issues in their planning? (2) Could this quietly become a gatekeeper rather than a convenience, if staff start expecting people to “talk to the bot first”?)
- Create a RAG framework for an issue of importance to you.
- A RAG tool over your major’s historical course evaluations that your peers can chat with (maybe “interview-me” style!) to help them choose electives, with citations back to original reviews. (Ethics reflections: Studies have shown that student course reviews suffer from systematic bias (e.g., by gender of the instructor) — could this tool damage careers of underrepresented faculty by further disadvantaging the success of their courses with students?)
- A RAG-based “explain this bill” tool over pending legislation or your city council’s agenda packets. (Ethics reflections: Could the tool’s framing of “what this bill does” introduce an unintentional slant?)
Examples using upcoming skills (or a mix of past and upcoming skills)
- Analyze a data set of importance to you, and make a dashboard that summarizes and visualizes important aspects of the data and/or set up the dataset as a RAG that you can interactively ask questions about through LLM chat interface. Need ideas for interesting datasets? Kaggle has tons, here are a few:
- Spotify music popularity (Ethics reflections: Is there something inherently flattening or dehumanizing in reducing art to a stats dashboard — or how could you design your dashboard to avoid that? If artists are meant to use the dashboard to predict how popular new songs-in-progress they’re working on will be, could that cause a harmful homogenization of music creation over time?)
- Gini coefficients (Ethics reflections: How could creating a tool for communicating this measure of income inequality to the general population help raise awareness and build support for policy interventions aimed at reducing inequality or mitigating its effects? On the other hand, could revealing the extent of inequality in stark terms dampen resolve to address it by making the problem seem too daunting?)
- Dating app reviews (Ethics reflections: Is there a danger that causing people to focus too much on how they “rate” on dating apps could lead to loss of self-esteem, or misguided focus on developing the self in shallow ways?)
- Create a multi-agent workflow where one model (perhaps an expensive larger and/or “thinking” one) generates some kind of content, and a second model (perhaps a smaller/cheaper one) checks the output to filter for some undesired quality that might have escaped the main model’s guardrail instructions.
- Build a small eval suite agent that comes equipped with test questions and a rubric that it automatically applies to output of other models. It would be used for comparing models across vendors and/or model sizes on a task you care about, and publish the results as a “report card.” (Ethics reflections: Evals can be unrepresentative or gamed — what would make yours biased, and how would you disclose that?)
- An “audit trail” agent layer for an AI workflow you already use: create an agent profile using a second (cheap) model that logs every claim generated by your main (larger) model and flags claims as verified/unverified/contradicts-source. For example, a tool like this could help check the work of lawyers who use AI to make sure there aren’t any hallucinated citations. (Ethics reflections: if the audit layer itself misses something, who’s accountable — you for making a not-100%-reliable tool, the human lawyer, or the model developers?)
- A “red-team your own prompt” AI agent whose only job is trying to jailbreak or misuse a template you built for something else, before you deploy it for real. (Ethics reflections: Although you might plan to only use your tool on your own work before it goes out to the public, if others gained access to your tool, could they use it for nefarious purposes? Even for your own use, there are questions about what to do with information about the weaknesses it might uncover, e.g., is it responsible to publish the weaknesses you find alongside the tool itself?)
- Use connectors to manage your personal productivity.
- A virtual “chief of staff” agent that does things like prepare daily morning briefings that summarize inbox and calendar and synthesizes a to-do list, perhaps even develops a first draft response to some emails. (Ethics reflections: How would your work associates, friends, and family feel if they knew that their interactions with you were being mediated through an LLM schedule manager, or that your email responses were being largely auto-generated? Over time, could you lose touch with your own life goals and priorities, becoming more of a passenger in the car of your life?)
These are just ideas I hope serve as some inspiration for what will be possible by the end of the quarter, not set a bar of expectation that all this should feel achievable to you at this moment (in fact, I expect much of it won’t — yet!). I genuinely look forward to seeing what fresh ideas you bring to life!