Every year, thousands of homebuyers close on properties without understanding the full risk picture — hidden title defects, wildfire exposure, flood zones, and earthquake proximity.
Title insurance companies spend millions manually underwriting these risks. We set out to build a tool that makes professional-grade property risk analysis accessible to any buyer or agent in seconds.

What it does

PROP.INTEL generates a comprehensive risk report for any U.S. property address in under 10 seconds.

It produces two scored risk dimensions:

Title Risk

Flags foreclosures, liens, rapid flips, and ownership chain anomalies using real county records.

Hazard Risk

Uses a machine learning model trained on FEMA National Risk Index data across 3,000+ U.S. counties.

Ridge regression fits hazard weights and min-max scalers on county-level Expected Annual Loss (EAL) figures
Produces a 0–100 composite score across 8 perils:
- Earthquake
- Wildfire
- Flood
- Hurricane
- Tornado
- Hail
- Strong Wind
- Coastal Flood

The model then estimates annual dollar losses for the specific property by scaling county EAL against the property's assessed value, generating per-peril figures such as:

~$2.3K/year wildfire exposure
~$890/year flood loss

AI Synthesis Layer

An AI summary (GPT-4o) synthesizes both dimensions — and explains why each hazard score is high or low based on real local geography — into a plain-English underwriting recommendation:

Proceed · Caution · High-Risk · Avoid

How we built it

Frontend

Next.js 14
TypeScript
Tailwind CSS
Recharts (interactive hazard breakdown visualization)

Backend

Next.js API routes
Server-side rendered property pages calling internal API endpoints

Property Data

ATTOM Data API (ownership history, sales, AVM estimates, property details)

Hazard Model

FEMA National Risk Index (2023) dataset loaded at build time
Ridge regression across 3,000+ U.S. counties
Min-max scaling to normalize Expected Annual Loss into a 0–100 composite score

AI Layer

GPT-4o-mini → location-specific hazard explanations grounded in FEMA EAL values
GPT-4o → final structured risk narrative

County Resolution

FCC Census Block API (primary)
GPT fallback for coordinate → county FIPS mapping

Challenges we ran into

ATTOM free tier capped at ~100 API calls/day
FEMA NRI is county-level; parcel-level approximation required building stock ratio adjustments
Passing live FEMA EAL data through four component layers to power dynamic Recharts tooltips
Inconsistent geocoding coverage required layered FCC + GPT fallback for reliable county resolution

Accomplishments we’re proud of

A real actuarial model backed by FEMA data — not synthetic scoring
Per-property annual loss estimates (e.g., ~$2.3K/year wildfire exposure) derived from county EAL scaled by AVM
AI explanations referencing real local geography
Full portfolio tracking for side-by-side property risk comparison

What we learned

Public datasets like FEMA NRI and FCC Census APIs become powerful when paired with ML normalization
LLMs work best as a synthesis layer on top of structured data — GPT explains the numbers, it does not invent them
Next.js API routes allow secure server-side key management while maintaining fast SSR performance