A full-stack rewrite of a legacy Flask/OpenCV school attendance prototype into a real-time, production-shaped application: React + Tailwind, FastAPI, and PostgreSQL + pgvector for both relational data and face-embedding search.
Project code lives in attendy-v2/ — see
attendy-v2/README.md for setup instructions and
attendy-v2/docs/adr/ for the reasoning behind the
harder decisions.
- Enrolls a student's face through a browser-guided burst-capture wizard — just a webcam, no special hardware.
- Recognizes faces live over WebSocket, with per-track identity assignment via the Hungarian algorithm, smoothing across frames (4-of-6 agreement) and gating on a motion-based liveness check before ever writing an attendance record.
- Marks attendance instantly in Postgres the moment a face is confirmed, and pushes that confirmation to every open dashboard tab in real time — no manual refresh.
- Filters the attendance sheet by class, section, date, and status entirely server-side, exports it to Excel, and surfaces attendance-rate trends and chronic-absentee flags on an analytics dashboard.
The original prototype used OpenCV's LBPH face recognizer trained on 1-3 low-resolution photos per student — with enrollment and live recognition even using different detection parameters, a real bug rather than a tuning problem. Rather than patch it, this project replaces the recognition engine (ArcFace embeddings + pgvector similarity search), the data layer (CSV files → normalized Postgres schema), and the frontend (server-rendered templates → React/Tailwind), while deliberately keeping the one part of the original design that was genuinely good: temporal smoothing across frames before ever trusting a recognition.
React · TypeScript · Tailwind CSS · Vite · FastAPI · SQLAlchemy (async) · PostgreSQL · pgvector · InsightFace (ArcFace) · WebSockets · Docker · GitHub Actions
cd attendy-v2
docker compose --profile full up --buildFull setup, testing, and deployment instructions: attendy-v2/README.md.