Inspiration

Every USC student knows the housing scramble. Rooms near campus get posted in a dozen group chats and spreadsheets, subleases come with no idea who you'd be living with, rent looks either too good or too high with nothing to compare it to, and scam listings target students who can't tour in person. Students leaving for internships or a semester abroad still owe rent until someone takes over their lease. We wanted one place where a Trojan can find the room, the roommates and the truth about the building.

What it does

HOMIE is a housing network for USC students, with sign-in limited to @usc.edu emails.

  • AI match from plain English. Type something like "private room under $1,400 within a 15 minute walk, quiet, vegetarian roommates, from January". An LLM turns it into structured preferences, our scorer ranks every listing, and the model re-ranks the top picks with a reason for each one, grounded only in the listing's real data.
  • Explainable match scores. Every room and roommate gets a score built from budget, dates, distance to campus and lifestyle (sleep schedule, cleanliness, guests, smoking, drinking, diet, pets, personality), and it always shows why.
  • Fair-price check on every listing. A badge like "$275 above similar" compares rent with the median of genuinely similar listings: the same kind and size of space, the same walking distance to campus and the same move-in timing. It always says exactly what it was compared against.
  • Red-flag detection. Listings are checked for bait prices far below market, pay-before-viewing language and text copied from other posts.
  • Roommate notes on subleases. The student leaving a room describes the roommates, so you know who you'd be living with.
  • Swipe to discover, a 3D map with walk times to campus, in-app messages, roommate reviews you can leave once you've talked with someone, and building intelligence that turns resident reviews into findings with cited quotes.

AI match: the request, what HOMIE understood, and ranked picks

Each pick comes with a grounded reason and who you'd live with

Fair-price check against genuinely similar listings

Roommate notes from the student leaving a sublease

3D map with walking rings to campus

How we built it

  • Backend: FastAPI, Pydantic v2 and SQLAlchemy 2 on PostgreSQL with pgvector, with a SQLite fallback for serverless.
  • AI: a provider-agnostic LLM and embeddings layer. In production it runs GPT-OSS 120B on Groq, with schema-validated structured outputs and a keyword-based fallback, so search still works if the model is rate-limited.
  • Matching: a deterministic, explainable scorer. The LLM only parses requests and explains its picks, receives the same walk times the site shows, and can only cite listings that exist.
  • Frontend: Next.js 16, React 19, TypeScript, Tailwind CSS v4, Motion, three.js and Leaflet, in USC cardinal and gold. Listings without photos get a generated room illustration that matches the kind of space.
  • Auth: passwordless 6-digit codes sent to @usc.edu addresses through an HTTPS email relay. Codes are stored hashed, expire after 10 minutes and are rate-limited.
  • Deployment: Vercel (Next.js and the FastAPI service in one project), with CI running lint and 98 backend tests on every push. The live site is seeded with about 3,000 fictional USC-area listings so the matching can be tested at real scale.

Challenges we ran into

  • Keeping the AI honest. We made the scorer the source of truth and limited the LLM to understanding requests and explaining results, dropping any listing ID it invents and giving it the exact figures shown on each card.
  • Free-tier rate limits. We cut prompt size, capped output tokens and added a retry and a non-AI fallback.
  • Serverless email. Vercel functions couldn't reach Gmail SMTP, so we built an HTTPS relay that only accepts @usc.edu recipients.
  • Defining a fair price. Comparing a room with a whole apartment is meaningless, so we classify each listing's kind and size of space and only widen the comparison when there isn't enough data.

Accomplishments that we're proud of

A working product, not a mockup: real sign-in, real AI search over thousands of listings, explainable scores, fair-price and scam checks, and a polished interface, built by four people in one weekend.

What we learned

Trust in an AI product comes from showing the evidence: sample sizes, the reasons behind a score, and exactly what a price was compared with.

What's next for HOMIE

  • Group house hunt: rank whole units for a group of friends, with per-person fit and a rent split.
  • Lease checker: upload a lease and get the clauses that matter explained in plain English.
  • Sublease rescue: alerts that connect students who need to leave a lease with seekers whose dates fit.

Slides: https://docs.google.com/presentation/d/1BZBlKmTDsgUBp7-wq4ZM9w9_LbCWebHi4gAvqRN_Iug/edit

Built With

Share this project:

Updates

Submission history