Pivot Argus from plate scanner to building intelligence by jtannahill · Pull Request #1 · jtannahill/argus

jtannahill · 2026-03-19T20:45:20Z

Summary

Complete pivot from license plate scanning to real-time building identification
Point camera at any NYC building → instant story, ownership, value, violations, permits
Two-phase ID: GPS + heading (<500ms) → CLIP visual matching (1-3s async)
8 NYC public data APIs (GeoClient, PLUTO, ACRIS, DOB, HPD, LPC, DOF, 311) — all free
Bedrock (Claude) generates building narratives
Free app monetized through aggregate scan analytics (heat maps, intent signals, imagery)

Changes

Backend (Python): New data model, NYC data provider, /identify endpoint, data assembly pipeline, Bedrock story generator, CLIP visual matching, scan analytics with heat counters
CDK (TypeScript): All 5 stacks updated — new Lambdas, Step Functions chain, CLIP SageMaker endpoint, S3 buckets for index/embeddings
iOS (Swift 6): BuildingIdentifier replaces PlateDetector, bottom sheet with 5 tabs, Explore map view, heading tracking
Dashboard (Next.js 16): Building heat map, building detail page, Mapbox dark style
ML: CLIP ViT-B/32 inference handler, Street View seeding script
Cleanup: Removed all plate-specific code (plates, geofences, alerts, pattern, diplomatic, old ML models)

Stats

91 files changed, +5,705 / -3,136 lines
127+ Python tests passing
34 CDK tests passing
Dashboard builds cleanly (3 static routes)

Test plan

Register NYC GeoClient API key
Package CLIP model to SageMaker
cdk deploy --all
Seed 5K Manhattan landmarks from Street View
Smoke test: point at building → story appears
TestFlight build

🤖 Generated with Claude Code

Replace plate-scanning key patterns (PLATE#, SIGHTING#, ENRICHMENT#, GEOFENCE#, CONNECTION#, ALERT#) with building-specific patterns (BLDG#, SCAN#, GEO#, HEAT#, ANALYTICS#, PROFILE, STORY, OWNER#, VIOLATION#, PERMIT#, IMAGE#). Add geohash6 encoder and 14 passing unit tests. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…PC, DOF integration Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

POST /identify receives {latitude, longitude, heading}, queries NYCDataProvider.resolve_location(), and checks DynamoDB cache for PROFILE and STORY records on each candidate BBL. Returns enriched candidates list with any cached data attached. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…ation Records individual scan events (SCAN#{userId}#{timestamp}/SCAN) and atomically increments daily building heat counters (HEAT#{date}/BLDG#{bbl}) via DynamoDB ADD. Scheduled aggregator rolls up yesterday's heat partition into an ANALYTICS#daily summary and flags trending buildings (scanCount >= 5). 13 tests, all passing. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Implements Task 4: NYCDataProvider-backed Step Functions task that assembles building data from all NYC public APIs (PLUTO, ACRIS, DOB, HPD, LPC, DOF) and writes normalized DynamoDB records (PROFILE, OWNER, VIOLATION, PERMIT). Adds Bedrock Claude story generator that reads the PROFILE and produces a headline, narrative, and fun facts, persisted as a STORY record. 23 tests, all passing. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- ml/clip_embedding/inference.py: SageMaker handler with model_fn/input_fn/predict_fn/output_fn for CLIP ViT-B/32, lazy-imports torch+clip for testability, returns 512-dim L2-normalized embeddings - ml/clip_embedding/requirements.txt: torch>=2.0, openai/CLIP, Pillow>=10.0, numpy>=1.24 - api/visual_match/handler.py: Lambda that fetches user photo from S3, invokes CLIP endpoint, computes cosine similarity against DynamoDB IMAGE# reference embeddings, returns best BBL match with confidence/method/corrected fields and stores new embedding in building index - Tests: 25 passing (10 inference, 15 handler) — all mocked, no GPU required Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…embly, CLIP, analytics) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Replace plate-centric UI with building intelligence dashboard. New home page shows Mapbox dark map with GeoJSON heat layer (green→yellow→red by scan density). /buildings?bbl=XXX shows tabbed detail: story, owner, value, violations, permits. Removes SearchBar, AlertsPanel, DiplomaticBadge, PlotMode, TimelineScrubber, and /plates route. api.ts rewritten with getBuilding, getRecentScans, getTrending, searchBuildings. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…fication Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

… diplomatic, ML models) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Adds ml/seed_index/seed_streetview.py — a CLI tool that queries NYC PLUTO via Socrata for Manhattan buildings, fetches 2-4 Street View images per building at configurable headings (default 0/90/180/270°), uploads each to S3, generates a CLIP embedding via SageMaker, and stores an indexed record in DynamoDB. Skips sub-5 KB Street View placeholder responses. Supports --dry-run, --limit, and full endpoint/bucket/table overrides. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>