rhconsole

Context

Recruiting means reading many CVs against many job offers. rhconsole automates the first pass: CVs and offers are dropped in watched folders or uploaded in the dashboard, and each pair gets a 0–100 score with the reasons behind it, so a recruiter knows what to check before calling a candidate.

What I built

• A FastAPI REST API with 40+ endpoints, JWT authentication with three roles, rate limiting and Prometheus metrics, on PostgreSQL 16 with 25 Alembic migrations.
• An event pipeline: a folder watcher and a standalone sync agent feed a Redis Streams queue, and a worker retries with exponential backoff before extraction, profiling, embeddings and scoring.
• Text extraction that reads two-column CV layouts column by column, with a per-page Tesseract OCR fallback for scanned PDFs, tuned through audits on 13,000+ real CVs.
• A four-layer skills taxonomy (2,162 hand-reviewed aliases, ROME 4.0, ESCO, O*NET and e-CF) with n-gram matching and a typo-tolerant fallback.
• A hybrid 0–100 score: skills, experience, education, languages and contract, plus a multilingual e5 embedding signal and a French cross-encoder. Coverage caps and a penalty for missing key tools stop a generic semantic match from hiding a skills gap.
• Every match comes with an explanation (why it matches, points of vigilance, evidence from the CV), three score levels and recruiter feedback stats, in a vanilla JavaScript dashboard.
• Docker Compose for local and cloud, a hardened API container (read-only filesystem, no Linux capabilities), and GitHub Actions CI that blocks a merge when a model changes without its migration.

Security

I audited the production app against OWASP practices and fixed real issues myself: a JWT secret left as an example value, a default administrator account exposed, and credentials present in the Git history.