Uplift ATS — CV Pipeline

Problem

A government HR client needed to process 124+ applicants for an OPF admin role with structured scoring against defined competencies. Manual review was creating bottlenecks and inconsistency.

Architecture & Solution

// System Pipeline

CV upload triggers OCR extraction → Claude Haiku evaluation against job description rubric (deferred async queue) → verdict written to Supabase → frontend polls status and renders distribution chart. Row-level security isolates tenants. Kill switch disables evaluation per-job without code deploy.

Implementation Decisions

1. Deferred evaluation: CV upload and AI scoring are decoupled — upload is synchronous, scoring is async to prevent timeout on large batches
2. OCR-first enforcement: Claude API is never called on a CV that has not cleared OCR — prevents token waste on blank/corrupt files
3. Dual-model pipeline: Haiku for CV parsing via OCR-first pipeline, Sonnet for structured scoring — LLM fires only after both CV and Excel form data are present
4. Excel master sheet importer: 39-field CNIC-normalized mapping, dual-source evaluation logic (form-only vs. form+CV)
5. Strict state machine: pending_data → ready_to_evaluate → evaluating → complete/error/ineligible — prevents premature AI calls and ensures audit-grade evaluation logs
6. Supabase Auth migration from Neon Auth/Better Auth (RS256→HS256) eliminated JWKS-related session bugs
7. Kill switch per job_id: lets ops pause AI scoring mid-batch without touching code

System Ownership & Proof

Gallery