Web / Done
ShelfOpsRetail Operations Platform
Overview
Retail operations OS that turns CCTV and shelf photos into actionable numbers: stable-ID visitor tracking, per-zone dwell, empty-shelf detection, price mismatch, stock and mutations, procurement, tasks, reports, and a Copilot that answers from SOP and store data with citations. Built as a full Docker stack: Next.js web, Go gateway and worker, Python AI services over gRPC, PostgreSQL 16 with pgvector, fronted by Caddy on a VPS.
Challenge
Unifying five layers (capture, understand, record, act, learn) into one readable system: CCTV and photo ingest, visitor and shelf detection, stock and price book records, alerts with SLA tasks, and a cited Copilot, all behind one domain with session cookies and a production guard that refuses to boot on mock AI providers.
Result
Shipped a healthy production stack with green healthz and readyz, passing end-to-end smoke tests (login, document upload, chat, audit to finalize, PO, idempotent GRN, IDOR guard), Playwright browser E2E on landing plus demo login flow, and a static demo mode that runs the web without any backend from gateway-shaped fixtures.
Key Statistics
Caddy, web, gateway, Postgres, ai-rag, ai-vision
Stack services
Login, chat, audit finalize, PO, GRN, IDOR guard
Smoke tests
Playwright landing desktop plus mobile
Browser E2E
Backend demo session plus static fixtures
Demo modes
Technologies
Frontend
-
Next.js
-
React
-
TypeScript
-
Tailwind CSS
- Radix UI
- Recharts
Backend
- Go
- chi
- pgx
- OpenAPI 3.1
- Protocol Buffers
- gRPC
AI
-
Python
-
FastAPI
-
ONNX Runtime
- RAG hybrid pgvector plus tsvector
Data
-
PostgreSQL 16
- pgvector
Deployment
-
Docker Compose
- Caddy
- GitHub Actions
Gallery
Store Summary
Operational overview with visitor counts, empty shelves, and active cameras for the demo store.
Inventory
Per-store and per-location stock balances with inbound and outbound mutation history.
Procurement
Purchase orders and goods receipts that add stock with an idempotent mutation trail.
Shelf Monitoring
Shelf photo audit flow from session to finalize, auto-creating alerts and tasks.
Tasks
Follow-up queue where OOS and price alerts become owned tasks with deadlines.
Copilot with Citations
Store-grounded Q&A over SOP and live data with sources and a daily cost cap.