AI Real Estate Lot Analysis Platform for a Residential Builder

The problem

A US residential builder buys land, not houses: the question for every listing is what can legally be built on it and which option makes money. Estimators were checking MLS listings, zoning tables and cost spreadsheets by hand, lot by lot. The engagement started in 2024 with a scraper, a valuation model and a pro forma engine; the current platform turns that into a product the whole team uses every morning, with an AI analyst on top and a redesigned public website.

What I built

Landing page. Live product widgets (ranked lots, top lots today) sit inside the hero, so the page sells what the tool actually does. The site was redesigned end to end with SEO and GEO in mind.

Dashboard. The morning view: county land inventory by zoning, every scored lot on the map, the AI's two best options with ROI and score, the team's shortlist and projected portfolio ROI. MLS sync status is always visible.

Search. The stored MLS feed with the fields an estimator actually uses: city, zoning, lot size, price, year built or vacant, unit type, days on market, plus a 22-field catalog (utilities at street, sewer vs septic, flood zone, slope, allowed unit types, max units, FAR). Sort by score, ROI or price; saved searches re-run after every sync.

Lot analysis. One lot, fully modelled: cost breakdown (land, hard, soft, margin), cottage vs duplex vs townhome ×4 vs multi-unit ×6 compared on all-in cost, ROI, margin and months, with "not by-right" flags where a permit is needed. A zoning card shows by-right uses, conditional uses and limits, and the AI's recommended option explains its reasoning. PDF and Word reports can be uploaded for the session only.

AI assistant. Ask in plain English, get a cited answer with the numbers. The analyst reads the shortlist, the live MLS database, the zoning code and any uploaded report, returns a comparison table and names its sources. Documents are discarded when the session ends and never used for training. The sync card shows the automated MLS update: listings stored, added, removed, price changes.

How it was delivered

The first phase (2024) was data and modelling: a Selenium bot captured 365 days of MLS history (active, pending and sold), pandas and SQLAlchemy harmonised it, an XGBoost automated valuation model priced homes, and a rule-based pro forma engine scored hundreds of lot-and-home-style combinations on profit, ROI and time-to-sell within zoning and quota constraints, handed over with documentation and reproducible code. The current platform replaces the manual loop with an automated MLS pipeline that pulls every land listing into Postgres twice a day, a Next.js application for the dashboard, search, lot analysis and assistant views, and an LLM analyst grounded in the shortlist, the MLS database and the zoning code that cites what it used. The public website was rebuilt at the same time.

Outcome

Estimators start the day with a ranked shortlist instead of a spreadsheet, see for any lot what can be built by right and which unit type pays, and can ask the analyst for a cited comparison. The client's Upwork review calls the result an outstanding job that exceeded expectations.