Project Brief

Built for Pollák Antal Technical School, Pollák E-Tabló solves a specific real-world problem: finding people in scanned yearbook images quickly and reliably. I designed it as a full platform, not just a model demo, with secure admin workflows, long-running processing support, and manual correction tools.

System Architecture

• Frontend (React + Vite + TypeScript): public browsing, admin upload/editor flows, and Keycloak login handling.
• Backend (Express + Prisma + MySQL): API orchestration, metadata persistence, protected admin endpoints, and yearbook/person CRUD.
• OCR service (FastAPI + OpenCV + RapidOCR): text extraction and YuNet-based face detection for candidate boxes.
• Infrastructure: Docker Compose orchestration for frontend, backend, and OCR microservice communication.

How I Built It

I built the pipeline around reliability for large image processing:

1. Chunked upload flow (/upload/init, /upload/chunk, /upload/complete) to avoid failures on large files.
2. Background processing session in the backend with progress state and ETA tracking.
3. Face detection + OCR extraction in the Python service, then structured candidate data returned to the backend.
4. Server-Sent Events progress stream so users see live status instead of waiting on blocking requests.
5. Manual correction editor with zoom/pan, box editing, add/remove person tools, and final persistence to MySQL.

Security and Access Model

Admin operations are protected with Keycloak token validation and role checks. The backend validates JWT issuer/signature and applies stricter limits on administrative endpoints. This keeps editing and upload features controlled while public search remains accessible.

Challenges and Engineering Decisions

• Large-image stability: Chunked upload and high server timeouts reduced failed requests during heavy processing.
• OCR quality on real data: I used PP-OCRv5 Latin recognition plus image preprocessing to improve name extraction quality.
• Duplicate/false face boxes: Non-max suppression and geometric filtering helped keep face results usable.
• User trust in long tasks: SSE progress and explicit step messaging improved the processing UX significantly.

Outcome

Pollák E-Tabló became a complete production-style workflow from raw image upload to searchable, corrected yearbook data. This project represents my strongest backend-heavy portfolio piece: distributed services, auth, async processing, and UX designed around technical constraints.