Type any San Francisco intersection. GPT-5.6 surveys street-level and north-up satellite imagery behind a blind-data firewall; crash, 311, district, and legislative records then corroborate the result. The real street is rebuilt as an explorable 3D miniature, one toggle applies priced grant-matched fixes, and an addressed resident letter opens in the correct Supervisor's email channel.

Built for OpenAI Build Week 2026, track Apps for Your Life.

Watch the 2:07 public demo

The product does not show an AI a scary crash history and ask it to find a matching problem. Stage 1 is firewalled: GPT-5.6 receives only four street-level images plus an optional north-up satellite view and must report only visible conditions in fixed street zones. Crash and 311 records are fetched separately and introduced afterward. The satellite view improves crossing-geometry coverage without revealing the location name or record history.

intersection search
    ├── Nominatim geocode
    ├── parallel: Street View + satellite · crashes · 311 · OSM · district · Legistar
    ├── LOOK   GPT-5.6 blind vision pass (images only)
    ├── CHECK  whole-word code matcher → CONFIRMED / CANDIDATE / REPORTED
    ├── FIX    deterministic cost + grant rules
    ├── ACT    GPT-5.6 addressed resident letter + social post
    └── Three.js real-street diorama → TODAY / PROPOSED

The server streams real stage completions over Server-Sent Events: counts appear only after the corresponding source or stage completes. The UI also exposes warnings rather than hiding partial-source failures.

• A point-in-polygon lookup against DataSF's current Supervisor Districts dataset resolves the district and current officeholder. Contact details are linked to the official Board roster, the letter names the Supervisor, and Open addressed email creates a populated mailto: draft.
• A district leaderboard groups reversed street pairs under one canonical intersection and ranks returned injury-crash records with a published score: 12× fatal + 5× severe + 2× pedestrian + crash count. In judge mode, non-demo corners are visibly disabled; live mode opens any ranked corner with one click.
• The legislative paper trail uses the official Legistar API and shows file number, recorded status, action, and date. The product never interprets an empty match as proof that no City action occurred.
• Export postcard captures the WebGL scene in TODAY and PROPOSED states and downloads a 1600×1000 share card with evidence counts, official, interventions, and planning cost.

The 3D scene is data-derived rather than decorative: OSM polylines become rounded road segments, real footprints become extruded pastel buildings, and mapped crossing/signal nodes become street furniture. Findings use stable named zones, so a finding can anchor to the model even when image dimensions change.

TODAY	PROPOSED

npm install
DEMO_MODE=1 npm run dev

Open http://localhost:5173. The app opens directly on a complete 16th Street & Mission Street run with real DataSF, district, legislative, and OSM records refreshed on July 21, 2026. The fixture deliberately omits licensed Street View and satellite pixels; its visual findings and advocacy copy are clearly disclosed as a representative cached judge-mode response. The API still exercises the full product shape, and the TODAY / PROPOSED toggle adds visible crosswalk, bulb-out, and protected-bike-lane geometry.

Requirements: Node.js 20+ and npm.

cp .env.example .env
# Set DEMO_MODE=0, then add OPENAI_API_KEY and GOOGLE_MAPS_KEY
npm install
npm run dev

The web app runs at http://localhost:5173; Vite proxies /api to Express on port 8787. For a production-style run:

npm run build
DEMO_MODE=1 npm start
# open http://localhost:8787

Never expose keys to the browser. Both OpenAI and Google calls are server-side, .env is ignored, and only .env.example is committed.

The exact current identifier was checked against the official OpenAI model guidance before implementation. The docs state that gpt-5.6 is the family alias routing to the flagship gpt-5.6-sol; the app intentionally uses the alias and keeps it configurable through OPENAI_MODEL. Both AI stages use the Responses API with low reasoning effort:

• Stage 1: four street-level inputs and optional north-up satellite context plus the blind site-survey prompt. No location name, crash record, complaint, or OSM tag is included.
• Stage 4: the already-corroborated findings, up to 25 crash and 25 311 record excerpts, up to 3 legislative matches, the record summary, fixes, costs, grants, civic contact, and location are supplied for the letter and social post. Stage 4 is not firewalled: it runs after corroboration and intentionally has full context to write a grounded letter.

Both model stages retry malformed JSON once with a narrow correction instruction, then fall back without crashing. Record-only evidence remains visible as REPORTED even when Street View or the blind pass is unavailable.

If either key is missing or an AI request fails, geometry and public records remain available, visual claims are omitted, and a deterministic grounded letter fallback is used. The UI never invents a successful AI result in live mode.

Official reference: Using GPT-5.6.

• Traffic crashes: DataSF Socrata dataset ubvf-ztfx, 100-meter within_circle(point, …) query, newest 50 records.
• District leaderboard: the same crash dataset, current Supervisor District field, intersection-only records since 2021, and canonicalized street-pair grouping.
• Civic routing: DataSF cqbw-m5m3 current district polygons with a local point-in-polygon lookup; contact data is paired with the official Board roster.
• Legislative records: Legistar's official sfgov API. Results are exact two-street matches and preserve official statuses without interpreting gaps.
• 311: DataSF dataset vw6y-z8j6, recent records in an approximately 100-meter latitude/longitude box, then a whole-word street-safety filter. The current dataset's numeric lat/long fields work reliably; the older within_circle(point, …) recipe currently returns a Socrata argument error.
• Street geometry: OpenStreetMap via Overpass, 95-meter road/building/furniture query, with overpass.kumi.systems as a fallback.
• Geocoding: Nominatim with a named User-Agent and San Francisco scoping.
• Imagery: Google Street View Static API metadata check followed by four 640×640 views, plus a north-up 640×640 Maps Static satellite image at zoom 20.

Public services can rate-limit or time out. Each source has a graceful empty/error state; OSM retries a second endpoint. For a reliable review, use judge mode first.

Matching is deterministic and tested. Terms must match at whole-word boundaries, preventing errors such as king matching parking. Pedestrian observations can match Vehicle/Pedestrian crash types and curb/crosswalk/sidewalk reports; visibility findings can match broadside/head-on crashes and sign/signal/visibility reports. At least one relevant official record makes a visual finding CONFIRMED; otherwise it stays CANDIDATE. Unmatched official records are preserved as REPORTED.

Costs are planning-level values for advocacy, not construction bids. A field engineering review remains necessary before implementation.

npm run lint
npm test
npm run build

Tests cover:

• projection displacement and coordinate orientation;
• whole-word corroboration and reported-only records;
• deterministic fix/cost/grant mapping;
• cache rejection of empty payloads;
• concurrent inferior payload protection;
• completeness of the keyless judge fixture.
• district point-in-polygon resolution and current-office contact routing;
• canonical street-pair grouping and severity-weighted leaderboard ranking;
• exact legislative matching and non-inaction disclosure;
• satellite request/prompt isolation and postcard planning-cost totals;
• real HTTP requests against the Express app: malformed-query rejection, per-IP rate limiting, and forwarding a client's SSE disconnect into the analyzer's abort signal.

The cache scores payload completeness, serializes writes per intersection, writes atomically, and will not replace a stronger existing payload with an empty or thinner one.

server/                  Express API and analysis pipeline
  pipeline.js            orchestrates the four stages
  vision.js              blind GPT-5.6 image pass
  datasf.js              crash/311 clients + district leaderboard
  civic.js               district polygon + current office resolver
  legislative.js         official Legistar paper trail
  corroborate.js         independent evidence matching
  fixes.js               intervention, cost, and grant rules
  osm.js                  Overpass client with fallback
  cache.js                non-empty, best-payload-wins JSON cache
  lastmile.js             GPT-5.6 letter plus grounded fallback
web/src/
  App.jsx                 search, evidence brief, findings, advocacy UI
  postcard.js            side-by-side 3D share-card exporter
  three/city.js           OSM-to-Three.js road/building renderer
  three/props.js          signals, zebra, RRFB, cushions, bulb-outs, bike lane
  three/orbit.js          hand-rolled pointer, wheel, and pinch camera
sample/
  16th-and-mission.json   complete no-key judge fixture
test/                     Node test suite

Concrete acceleration moments from the build session:

• verified the live GPT-5.6 alias in official OpenAI guidance before wiring the Responses API, avoiding a speculative model slug;
• found and corrected the current DataSF 311 spatial-query mismatch while preserving the intended 100-meter behavior;
• generated the OSM projection, footprint extrusion, real crossing furniture, hand-rolled orbit controls, and remodel prop layer as one cohesive Three.js system;
• turned the known empty-response cache race into a scored, atomic best-payload-wins cache with regression tests;
• exercised the actual browser scene in both TODAY and PROPOSED states, not only the unit layer.

The remodel reveal is deliberately staged: improvements pop into the miniature 350 ms apart with their intervention, cost, and grant labels, while fatal-crash locations shift from a restrained red base rim to green in the proposed state.

All code here is original and built during OpenAI Build Week. The OSM-derived 3D remodel layer is original to Model Citizen.

OpenStreetMap data is © OpenStreetMap contributors, ODbL. DataSF records are provided by the City and County of San Francisco. Street View content remains subject to Google's terms and is not embedded in the public fixture.

Released under the MIT License.