Working with us (if you have a non-traditional background)
Interested in working with our group, but coming from a non-traditional background, e.g., industry, a field outside CS, or high school? You don’t need a CS degree or research experience. If you’re a CMU student or prospective postdoc, reach out to us directly instead. Or, if you are a prospective PhD student, apply to CMU CSD.
Here’s a taste of the kinds of problems we work on. One of the systems we are building is Quail, a SQL query engine with AI functions: you write prompts inside a SQL query, and Quail runs them over your data with open-source LLMs.
To get started, try Quail yourself: pick a dataset (see the ideas below) and analyze it with Quail. The open-source repo has documentation, and you can run Quail on Modal, whose free plan includes $30 of compute per month (see the Modal quickstart).
Then, email shreyashankar@cmu.edu a short video (under 10 minutes) of you explaining your analysis, along with your CV. Feel free to present it with a dashboard, slides, or another visual tool. You don’t need to share your code.
What makes a project good
A good project does at least one of these:
- Applies Quail to an interesting or impactful problem.
- Shows a kind of query people need that Quail can’t run yet, and how far you get with what Quail can do today.
- Identifies ways to make Quail faster, cheaper, or more accurate.
- Finds a better way to write Quail queries, or to communicate their results to people, e.g., an interface, a dashboard, or a story.
Ideas
These are just some ideas for inspiration. If you have your own data or questions, explore those! Pick something you’re interested in, or a domain you already know well.
Federal rulemaking. Regulations.gov has public comments on proposed rules, sometimes hundreds of thousands per rule. What do people argue for and against? How much comes from organized campaigns? Do final rules respond to the most common arguments?
Consumer finance. The CFPB complaint database has millions of complaints about banks, lenders, and debt collectors, many with the consumer’s own description. Which practices hurt people the most? Which ones show up across many companies? What changed after new rules?
Vehicle safety. NHTSA publishes owner complaints about vehicles going back decades. Which defects showed up in complaints long before a recall? Which problems are common now but haven’t been recalled?
Courts. CourtListener has millions of court opinions. How are courts responding to lawyers who cite cases that AI made up? How do outcomes on the same legal issue differ across courts?
State legislation. Open States has bills from all 50 state legislatures. What are states doing about AI, housing, or data privacy? Which bills copy each other, and where does model legislation come from?
Clinical trials. ClinicalTrials.gov has close to 600,000 registered studies. Which trials could a given patient join? Why do trials stop early? Who gets excluded from trials, and for what reasons?
Corporate disclosures. EDGAR-CORPUS has the annual reports of all US public companies from 1993 to 2020. How have companies described climate, cyber, or supply chain risks over time? Which risks did companies warn about before a crisis hit?
History. American Stories has articles from nearly 20 million pages of American newspapers from 1774 to 1963. How did coverage of an event, like the 1918 flu, spread across the country? How did papers in different regions describe the same events?
AI use. WildChat has millions of real conversations with chatbots, and Nebius has 80,036 coding agent trajectories. What do people actually use chatbots for, and where do chatbots fail them? Why do coding agents fail?