ML / DL
Store Intelligence
Turns raw CCTV footage into structured behavioural events and live retail analytics — an RT-DETR + ByteTrack vision pipeline behind a containerised FastAPI service.
01 / Problem
The problem
Retail stores record hours of CCTV that nobody watches. The footage holds real signal about dwell time, footfall, and queue formation, but none of it is queryable.
02 / Response
The solution
Built an end-to-end pipeline that ingests raw camera clips, detects and tracks people with RT-DETR and ByteTrack, emits structured behavioural events into Postgres, and serves them through a FastAPI analytics layer with a live terminal dashboard.
03 / Product
Key capabilities
RT-DETR detection with ByteTrack multi-object tracking
Structured behavioural event extraction
FastAPI + Pydantic v2 analytics service
Postgres event store
Live terminal monitoring dashboard
Full Docker Compose stack
04 / Outcome
Impact
05 / Reflection
What I learned
Tracking quality dominates downstream analytics — a small ID-switch rate in the tracker quietly corrupts every dwell-time metric computed on top of it.