- of HR and recruitment workflows automated
- ~30%
- of HR and recruitment workflows automated
01 — Problem
The problem
Keyword filters mismatch candidates and jobs: strong candidates who describe their skills differently get filtered out, and recruiters spend their time reading CVs by hand.
Any automation had to be explainable and auditable. A black-box score was not acceptable for hiring decisions.
02 — Role
My role
- Designed the extraction-and-matching pipeline and the human review workflow.
- Built structured CV extraction with Vertex AI LLMs, validated against controlled skill vocabularies.
- Delivered the React recruiter dashboard and the PostgreSQL audit trail.
03 — Architecture
How it fits together
01 Input
- CV upload
- Job description
02 Extraction
- LLM → JSON schema
- Skill vocabulary validation
03 Matching
- Embedding similarity
- Rubric scoring
- Quoted evidence
04 Review
- Recruiter shortlist
- Audit log (PostgreSQL)
Key decisions
- Schema-constrained extraction, so the model fills fields we control instead of writing free text.
- Every match explanation quotes the CV, so a recruiter can check the claim in seconds.
- Names and photos are hidden during first-pass review where policy allows.
- Pass rates are compared across groups periodically to catch drift.
04 — Results
Results
- ~30% of HR and recruitment workflows automated
- Recruiters review a ranked shortlist with reasons instead of an unsorted inbox
05 — Retrospective
What I’d do differently
Bring legal and HR into rubric design from the first sprint, not at the review stage, because their input shapes the scoring model.
Version prompts and extraction schemas like code, with regression tests on a fixed CV set, because small prompt edits can shift scores.
Working on something similar?
I’m happy to talk through architecture, trade-offs or a role.
Get in touch