LeaseLeak

Drop your rent roll. Five seconds later: how much rent you're under the HUD benchmark, which leases end in the wrong month, and the renewal letters — written.

Built solo by Rishik Rontala for VentureFix 2026 (Venture Build track). Live: https://leaseleak.vercel.app/ (mirror: https://rishikrrontala-bot.github.io/leaseleak/)

What it does

1. Match — every unit in your CSV/XLSX is matched to HUD's Small Area Fair Market Rent for its ZIP code and bedroom count (FY2027; 38,601 ZIPs, studio–4+ BR).
2. Measure — gap = max(0, SAFMR − rent) × 12, summed across the portfolio. FMR is the 40th percentile of local gross rents, so the number is a conservative floor.
3. Time — Zillow's ZORI rent index by ZIP (Jul 2026) becomes a 12-month seasonal curve (per ZIP → metro → national fallback). timing = rent × (idx_peak ÷ idx_lease_end − 1) × 12. The renewal term (6–18 months) is chosen so the next lease ends at the seasonal peak.
4. Write — one renewal letter per unit: new rent = min(SAFMR, rent × (1 + cap)), the term, notice language. Download all as PDF.
5. Value — the gap priced as building equity: value = Σ gap ÷ cap rate. On the sample portfolio $29,280/yr is ~$488,000 of value at a 6% cap. Cap rate is adjustable (4–10%).
6. Vouchers — HUD FMR is also the basis for Housing Choice Voucher payment standards. Units below the standard (FMR × 90–110%, set per housing authority) would earn max(0, standard − rent) × 12 more from a voucher household — a way to close the gap at turnover without raising rent on a sitting tenant.
7. Twelve months — projected monthly income under three policies (do nothing / proposed renewals / every renewal at benchmark), each increase applied the month after its lease ends.
8. Building grades — one A–F grade per property on gap ÷ gross rent (A < 2%, B < 5%, C < 9%, D < 14%), sorted worst-first with $/door and a one-line diagnosis.
9. Momentum — the ZIP's ZORI year-over-year move classified as rising (≥ +3%), steady, or softening (≤ −1%), with renewal advice.
10. Concentration — share of monthly rent whose leases end in the same month, the loss if the two largest don't renew, and a staggered calendar across the above-average seasonal months when it lowers the peak.
11. Utilities — FMR is a gross rent; marking a building "we pay utilities" adds a typical allowance by bedroom count ($95 / 120 / 150 / 180 / 210) before comparing.
12. Fairness — each proposed rent as a share of the ZIP's median household income (Census ACS 5-year); above 30% is flagged in the letters list and summarised before download.
13. If they leave — every increase priced with turnover risk: stay = increase × 12; leave = (benchmark − rent) × 12 − (rent × months vacant + make-ready); expected gain at an adjustable leave rate, payback months per unit, and the break-even leave rate.

Everything runs in the browser. No upload, no account.

The AI layer

Every dollar figure on the page is computed in the browser by a deterministic engine from public data. The model does two things code can't:

• Writes the plan. POST /api/plan sends the engine's full output (unit figures only — tenant names are stripped client-side) to Gemini with a JSON schema and gets back a headline, 3–5 prioritised actions with the units and figures behind each, a "skip" list, and one caution. POST /api/ask answers free-text questions the same way. The prompt forbids new numbers; a client-side check (verify() in src/lib/ai.ts) then extracts every $/%/month figure from the reply and matches it against the engine's numbers, so the UI can say "23 of 24 figures trace to the engine" and flag the one that doesn't.
• Reads exports the parser can't. When the heuristic column detection fails, POST /api/map maps headers to the schema from six sample rows. Missing ZIPs are looked up from street addresses with the U.S. Census geocoder (exact, keyless); only addresses it can't match go to the model, and those are marked "?" in the UI.

The key lives in a Vercel function (api/_gemini.ts): origin allow-list, per-IP rate limit, 200 KB body cap, retries on 429/503, model gemini-3.6-flash with a small thinking budget for the plan and none for the fast paths. The GitHub Pages mirror calls the same functions cross-origin.

Run it

npm install
npm run dev

public/data/safmr.json and public/data/zori.json are built from the raw HUD and Zillow files with scripts/build-data.py.

Data & licences

• HUD User, FY2027 Small Area Fair Market Rents — public domain. https://www.huduser.gov/portal/datasets/fmr/smallarea/index.html
• Zillow Research, ZORI (ZIP level, smoothed, all homes) — © Zillow, free for non-commercial use with attribution. https://www.zillow.com/research/data/
• U.S. Census Bureau, ACS 5-year 2019–2023, table B19013 (median household income by ZCTA) — public domain. Built from the table-based summary file, no API key needed. https://www2.census.gov/programs-surveys/acs/summary_file/2023/table-based-SF/
• Photographs: CC BY-SA 4.0 / CC0 via Wikimedia Commons — see public/photos/ATTRIBUTION.md.
• Satoshi typeface — Indian Type Foundry, via Fontshare (free licence).
• Instrument Serif — Rodrigo Fuenzalida & Jordan Egstad, via Google Fonts (SIL OFL).
• Stack: React 19, Vite, Tailwind v4, PapaParse, SheetJS, jsPDF.

Not legal or financial advice. Check your state's notice and rent-increase rules before sending letters.