Captured from the live deployment, not a mockup
01 · The problem
Ranked 100,000 profiles in under 18 seconds on one CPU core with zero network calls, after proving per-profile LLM scoring infeasible at that scale and cost.
02 · How it works
- 01
Filtered 100% of honeypot records deterministically ahead of scoring, and generated auditable per-candidate justifications with trust scores.
- 02
Established first that per-profile LLM scoring was operationally infeasible at 100,000 candidates on cost and latency, then built the ranking to run in under 18 seconds on a single CPU core with zero network calls.
- 03
Scored profiles across four weighted dimensions, role fit and skills at 30% each, experience and logistics at 20%, behind a deterministic filter that removed 100% of honeypot records before ranking ran.
03 · What it cost, and what it returned
Generated auditable, factual justifications per candidate with trust scores, so a recruiter sees why someone ranked where they did rather than a bare ordering.