Uplift ATS — CV Pipeline
An AI-powered candidate evaluation pipeline — engineered for a government HR vendor to automate structured scoring and eliminate manual screening bottlenecks.
Problem
A government HR client was overwhelmed by manual screening, needing to process hundreds of applicants for an OPF admin role with rigorous, structured scoring against defined competencies. The challenge: human review was subjective, slow, and created massive bottlenecks in the hiring pipeline. They required an automated, audit-grade system capable of parsing complex CVs and mapping them directly to strict evaluation rubrics without hallucination.
Architecture & Solution
The pipeline is built around a decoupled asynchronous evaluation engine. Resumes are uploaded and synchronously passed through an OCR extraction layer. The extracted text is then queued for evaluation, where Claude maps the candidate's experience against a predefined job rubric. The resulting verdict is written to an isolated Supabase tenant, while the React frontend polls the status to render distribution charts in real-time.
Implementation Decisions
- Deferred async evaluation — Decoupling CV upload from AI scoring prevents browser timeouts on large applicant batches and ensures reliable processing.
- OCR-first enforcement — LLMs are only triggered after a CV clears the OCR stage, preventing wasted API tokens on blank, scanned, or corrupt PDF files.
- Dual-model pipeline — A lighter model handles structural parsing via the OCR pipeline, while a reasoning model executes complex, multi-variable structured scoring against the master rubric.
- Per-job kill switches — Operational controls allow administrators to pause AI scoring mid-batch without requiring a code deployment, ensuring manual override capabilities.
