Canonical version: rantum.xyz
Predictive real-estate intelligence that fuses fragmented municipal permit records with multimodal AI to score teardown probability, flippability, and off-market opportunity.
The Challenge
The most valuable real-estate signals are buried in the messiest data. Building permits, zoning changes, and code-enforcement records are filed across dozens of municipal systems — each with its own format, portal, and lag — and almost never connected to live listing data. By the time an off-market teardown or value-add opportunity shows up in a normal feed, the developers who watch permits already know.
The hard part isn’t the model; it’s making the underlying data usable at all: fragmented, civic, unstructured, and constantly changing.
What We Built
AddressIntel is a live real-estate intelligence platform covering the San Francisco Peninsula and Nantucket Island. A Python ingestion fleet, orchestrated on GitHub Actions, continuously pulls municipal permits and public listing data; the data is normalized, fused, and scored, then served through a B2B GraphQL API.
It runs on a hybrid public-benefit model — a free public transparency layer over neighborhood development trends, funded by a monetized enterprise API for developers and institutional investors.
At a glance: 10K+ sales records ingested, 3 predictive models in production, 2 live markets.
Predictive Scoring & Multimodal AI
Teardown & flippability models — proprietary probability scores for teardown, flippability, and bidding-war likelihood, trained on fused permit + sales history. The permit data is the signal incumbents don’t have: a demolition permit filed six months before a property hits the market is invisible to anyone watching standard listing feeds.
Vision-based condition scoring — Google Gemini reads listing photos to score interior and exterior property condition automatically, turning unstructured images into a model feature. No manual tagging, no scraped review data — just the photos that already exist on every listing.
Permit alpha — joining civic permit filings to listings surfaces off-market developer activity before it becomes visible in standard feeds. The latency between a permit filing and public awareness is weeks to months. That gap is the product.
Architecture
The stack is built to be cheap to operate and easy to extend into new markets:
- Pipeline: Python ingestion fleet on GitHub Actions cron, normalizing multi-source civic and listing data, with Gemini handling unstructured vision and NLP tasks
- Backend: Firebase Data Connect (GraphQL over Cloud SQL Postgres) with Firebase Auth; Stripe-metered access for enterprise clients
- Frontend: Next.js (App Router) + React on Firebase App Hosting, with Sentry edge error tracking
GitHub Actions as the orchestration layer keeps infra costs near-zero for a market-coverage product where freshness matters more than sub-second latency.
What It Shows
Turning messy, fragmented public records into a predictive product — and crossing domains, from data science into real estate, with point-in-time ML discipline so the scores reflect what was knowable at the time, not hindsight.
It’s the same through-line as ClearTrace: find the signal in data others write off as too messy to use, ship it as infrastructure. There it’s on-chain trace data. Here it’s municipal permit filings. The skill transfers.
What It Proves
That you can find signal in a data-rich space incumbents underuse, and turn it into a monetizable product — pipeline, models, API, and billing included. The same capability transfers to any business sitting on fragmented or underused data: marketplaces, fintech, insurance, logistics.
The question is never whether the data exists. It’s whether anyone has built the infrastructure to make it usable.
Andrew Maury is the founder of Rantum, a senior data science & ML studio. He previously scaled data infrastructure at 0x Labs and contributed to Dune’s open-source Spellbook. AddressIntel is live at addressintel.co.
Top comments (0)