Due Intelligence

Agentic, multi-source due diligence for NYC development sites

First place out of 62 teams — Vercel x Google DeepMind “Zero to Agent” Hackathon NYC. Built by Ashkan Radnia and U Jin Seah, with the event hosted in partnership with Cerebral Valley.

Due Intelligence turns an NYC address into a source-grounded site briefing. It brings zoning, development capacity, regulatory constraints, market context, and visual site conditions into one workflow so acquisitions teams can screen parcels in minutes instead of assembling the same facts across disconnected portals.

Watch the product demo →

What it does

• Screens a parcel from one address. NYC GeoSearch resolves the property, then the app pulls lot, zoning, flood, geometry, and incentive data in parallel.
• Keeps facts separate from AI judgment. FAR, buildable area, development scenarios, and constraint scoring are calculated deterministically before Gemini interprets the result.
• Adds visual site intelligence. Gemini analyzes Street View and satellite imagery for existing conditions, lot coverage, shadows, neighborhood context, and buildable-envelope cues.
• Compares opportunities side by side. Users can pin, sort, map, and compare multiple parcels across zoning, development upside, constraints, and AI-assisted visual signals.
• Turns analysis into a deliverable. A tool-using analyst answers follow-up questions with streamed UI components, while an isolated Vercel Sandbox generates downloadable HTML reports.

System flow

flowchart LR
    A["NYC address"] --> B["GeoSearch: BBL + coordinates"]
    B --> C["Parcel data pipeline"]
    B --> D["Vision pipeline"]
    C --> E["Deterministic zoning + scenario calculations"]
    C --> F["Gemini structured analysis"]
    D --> G["Street View + satellite reasoning"]
    E --> H["Map, report, and parcel comparison"]
    F --> H
    G --> H
    H --> I["Agent chat + sandboxed HTML export"]

Loading

Engineering highlights

Area	Implementation
Grounded agents	A Vercel AI SDK tool loop queries nine property, risk, market, and demographic tools instead of asking the model to invent missing context.
Reliable computation	Typed, testable functions handle FAR selection, development scoring, buildable square footage, and MIH scenarios outside the model.
Multimodal analysis	Parallel Gemini calls evaluate street-level and aerial imagery, returning structured assessments plus annotated visual artifacts.
Resilient data fetching	PLUTO is treated as the required source; FEMA, DCP GIS, MIH, geometry, and AI interpretation degrade independently when unavailable.
Generative delivery	JSON Patch streams structured analyst responses into real UI components. Report generation runs the Gemini CLI inside an isolated Vercel Sandbox with validation and compact-prompt retry behavior.

Data sources

Source	Used for
NYC GeoSearch, PLUTO, and MapPLUTO	Address resolution, parcel geometry, land use, building facts, zoning, and FAR
NYC DOB, HPD, and Open Data	Permits, violations, complaints, rolling sales, and nearby incidents
DCP GIS and FEMA NFHL	Zoning overlays, MIH areas, and flood risk
US Census ACS and HUD	Neighborhood income and fair-market-rent context
Google Maps imagery	Street View and satellite inputs for multimodal analysis

Every output is designed as a screening aid, not a substitute for zoning counsel or formal due diligence.

Stack

• Next.js 16, React 19, TypeScript, Tailwind CSS, and Turborepo
• Vercel AI SDK with Gemini Pro, Flash, and image models
• Deck.gl, MapLibre, and react-map-gl for the spatial interface
• JSON Render for streamed generative UI
• Vercel Sandbox for isolated report generation

Run locally

Requirements: Node.js 20+ and pnpm 9.15.9.

1. Clone and install dependencies.

   git clone https://github.com/ashkanrdn/DueIntelligence.git
   cd DueIntelligence
   pnpm install

2. Create the web app environment file.

   cp apps/web/.env.example apps/web/.env.local

3. Add the keys for the features you want to run.

   Variable	Purpose
   GOOGLE_MAPS_API_KEY	Street View and satellite analysis
   CENSUS_API_KEY	Census enrichment in analyst chat
   HUD_API_TOKEN	Fair-market-rent enrichment in analyst chat
   GOOGLE_GENERATIVE_AI_API_KEY	Optional server-side Gemini fallback; users can instead provide a key in the app
   NYC_OPENDATA_TOKEN	Optional higher-rate access to NYC Open Data

4. Start the app.

   pnpm dev

5. Open http://localhost:3000, add a Gemini API key through the key icon, and search for an NYC address.

Recognition

• First place: Zero to Agent: Vercel x Google DeepMind Hackathon NYC
• Winner announcement: project overview and result
• Demo: 03.21.26 Hackathon walkthrough

Project status

This repository contains the working hackathon prototype. The core parcel, comparison, vision, chat, and report-generation paths are implemented. Data availability and model output can vary by address; all zoning conclusions should be verified against primary records and qualified professionals.