Build (self-applied) · 6 weeks (Forge-shaped engagement, internal)
Internal — Talpro Universe · HR-staffing platform
Time-to-shortlist
9h → 38m
per role, typical volume
Drift detection lag
48h
vs. 60–90d prior
Recruiter agreement
0.81
vs. 0.62 with vendor #2
01 · Problem
Recruiter time-to-first-shortlist on Talpro's core CV pipeline was 9+ hours per role. Growth was bottlenecking on people, not software. Two commercial vendors had been trialled and dropped — both had silent drift within 90 days.
02 · Approach
Rebuilt the screening layer on our own stack: TALPRO-IQ 8-axis model matrix for provider selection, two-stage ranking with a small distilled model for the coarse pass and a larger model with retrieval for the top-100 refinement. Every inference logged, six evals wired to CI before cutover.
03 · Outcome
Time-to-first-shortlist collapsed from 9 hours to 38 minutes on typical volumes. Drift eval catches model degradation within 48 hours (vs. the 60–90 days the vendors took to notice). Stack runs on in-region infra with no raw PII leaving the perimeter.
“Once the eval harness was in CI, the argument stopped being 'is AI good enough'. It became 'what's the next eval we're not measuring'. That's the right argument.”
Your engagement starts with a Prism day.
One day. Your stack, your team, your constraints. Out with a shortlist, a decision log, and a charter you can defend to your board.
Book a Prism dayNext case
Bangalore GCC captive · Charter-phase · client anonymised