Cloud-hosted shoplift detection SaaS for retail stores. Ships YOLO11-pose behavior detection, RAG + VLM verification, and anonymized cross-customer learning — all without requiring customers to buy or install new cameras or servers.
Архитектур (2026-04-21): Centralized SaaS default product. Харилцагчийн камерууд өөрчлөгдөхгүй; Chipmo-ийн WireGuard VPN appliance-аар RTSP sub-stream-ийг төв сервер рүү татаж inference хийнэ. No recording by default (≤10s confirmed alert clip only). 3-phase infrastructure: Railway → Cloud GPU → Owned GPU (1000+ cameras). On-prem SKU available as premium option for data-sovereignty customers. See
docs/decisions/2026-04-21-centralized-saas-no-customer-hardware.md.
- Full architecture:
docs/01-ARCHITECTURE.md - Roadmap:
docs/02-ROADMAP.md - Tech specs:
docs/03-TECH-SPECS.md - Infra strategy (3-phase):
docs/04-INFRASTRUCTURE-STRATEGY.md - Customer onboarding:
docs/05-ONBOARDING-PLAYBOOK.md - Privacy + legal:
docs/09-PRIVACY-LEGAL.md - Pricing (internal):
docs/10-PRICING-BUSINESS.md - Task board:
docs/TASKS.md - All docs index:
docs/README.md
- Backend: FastAPI (async), PostgreSQL + TimescaleDB, Redis Streams
- AI: YOLO11-pose, ByteTrack, OSNet Re-ID, Qwen2.5-VL via vLLM
- Vector DB: Qdrant (per-tenant isolated + shared behavior taxonomy)
- Ingress: WireGuard VPN appliance (GL.iNet or Raspberry Pi)
- Central GPU: Phase A cloud serverless → Phase B rented RTX 4090 → Phase C owned RTX 5090 / L40S
- Frontend: React 19 + Vite + TailwindCSS v4
- Observability: Prometheus + Grafana + Loki
- Orchestration: Docker Compose → K3s (Phase C)
- Hardware-free onboarding — no camera changes, no customer server.
- Self-improving via shared taxonomy — cross-customer learning without exposing any tenant's PII.
- Privacy-first — no recording by default, per-tenant isolation, DPIA compliance with Mongolian data protection law.
- Cost-effective at scale — economies of shared GPU increase from Phase A to Phase C.
.
├── shoplift_detector/ # Main FastAPI backend
│ └── app/
│ ├── api/ # REST endpoints
│ ├── ai/ # Detection pipeline
│ ├── db/ # SQLAlchemy models
│ └── services/ # Business logic
├── security-web/ # React dashboard
├── alembic/ # DB migrations
├── observability/ # Prometheus + Grafana config
├── docs/ # All development documentation
│ ├── decisions/ # ADRs
│ ├── audits/ # Audit reports
│ └── spikes/ # Technical spikes
└── tests/ # Test suite
Development setup instructions: see CLAUDE.md and
docs/00-AS-IS-INVENTORY.md.
Production deployment (Phase A, Railway):
# Push to Railway project chipmo-prod
# Services: api, worker, web, postgres, redis, qdrant, wireguardOnboarding a customer: see docs/05-ONBOARDING-PLAYBOOK.md.
- Engineering phase 0 (rule-based pilot): shipped
- Engineering phase 1 (quick wins + VPN onboarding): in progress
- Infra phase A (Railway): active
See docs/TASKS.md for detailed task tracking.
Updated: 2026-04-21