The workflow before#
Point72 had 10 years of recruiting history: resumes in PDF, Word and scanned form, search firm assessments, interview notes and hiring outcomes. It sat in separate systems, and almost none of it could be searched or compared. When a new candidate came in, recruiters and analysts read the documents one by one and judged them against criteria that varied from team to team. Hours went into each first pass. The firm had a record of which past hires had worked out, but no way to learn from it, so every search started close to zero.
What we built#
We built a recruiting application on top of a new HR data lake. It brings every resume, assessment and note into one place, reads them, and gives recruiters a searchable history and a ranked shortlist with the reasons behind each ranking. Recruiters still make the call. They just start from an informed first pass instead of a stack of PDFs.
Ten years of history in one place
The data lake pulls documents from the applicant tracking system, HRIS and document stores, runs scanned files through OCR and links every record for the same candidate across sources. A recruiter can search a decade of talent data in one query instead of asking three systems.
Profiles you can compare
Topic models turn free text into comparable profiles, and classifiers sort and clean documents by type and quality. Qualifications and experience are extracted automatically, so two candidates described in very different language can sit side by side.
A ranking that shows its reasons
A scoring model trained on past hiring outcomes ranks candidates against shared benchmarks. Next to each score, the recruiter sees which factors pushed it up or down. We chose a simple, explainable model over a more opaque one, because recruiters won't act on a score they can't question.
Reviewers keep the models honest
Reviewers validate extractions and label edge cases in a dedicated review queue, and the models retrain as new hiring outcomes arrive. Human judgment goes back into the system instead of staying in someone's inbox.
How we delivered it
The work ran about 12 months, from architecture through phased rollout. We built the data lake and history search first, since that gave recruiters value before any model was in place. Scoring followed, tested against past outcomes, then reached recruiting teams in phases so each group could check the rankings against its own judgment before relying on them.
Before and after#
| Measure | Before | After |
|---|---|---|
| First-pass document review | Hours per candidate | Minutes per candidate |
| Historical talent data | Free text in separate systems | 10 years in one searchable store |
| Candidate ranking | Manual, varied by team | Scored against shared benchmarks, with reasons |
What's next#
The same data lake and model pipeline can take on more of the screen. An agent could read each new application end to end, draft a candidate summary and ranking, and send only borderline cases to a recruiter, so the team spends its time on the decisions that are actually close.
