AI CV Screening and Evaluation

MII E-Recruitment: ~45% failures in the original 22-pair screening run; no recorded rerun. AI supplies evidence; the recruiter makes the final decision.

~45%
Failure rate, original 22-pair screening run; no recorded rerun
0.13%
Parse failure rate in the later baseline (1 of 800 calls)
270+
Test files added from a zero-test baseline, with evaluation gates blocking deploys

Role: Creator and lead engineer, responsible for evaluation decisions
Period: Jun 2026 to Oct 2026
Project type: Featured
Room in the 3D HQ: L3:ai-cv-screening-platform

Results

~45%
Failure rate, original 22-pair screening run; no recorded rerun
Observed in a 22-pair evaluation run, with fixes covered by 7 regression tests from real error strings. A rerun of that exact evaluation was not recorded.
0.13%
Parse failure rate in the later baseline (1 of 800 calls)
Measured in a later evaluation baseline using a newer model and prompt, not a before/after estimate of the original fixes.
270+
Test files added from a zero-test baseline, with evaluation gates blocking deploys
Based on repository history.
21.1 s to 12.1 s
Screening p95 latency in controlled evaluations, medium vs high reasoning effort
Measured in controlled evaluation rounds; the selected setting still passed fairness and injection gates.

Problem

LLM screening frequently failed and issued hire-or-reject verdicts without evidence. There were no tests or release gates to catch regressions from prompt or model changes.

Contribution

I traced two production errors to token exhaustion mid-JSON and raw newlines in strings, then addressed them with JSON mode, output caps, a quote-aware extractor and retries.
I introduced the first test suite and CI-gated evaluation harness, using a 40-pair golden set (37.5% in Bahasa Indonesia), counterfactual fairness checks and 20+ prompt-injection classes.
I removed AI hire-or-reject verdicts and added AI disclosure, a system card and identity redaction before model calls. In October 2026 I redesigned every page in English and Indonesian: the candidate result now leads with a fit band, requirement coverage and must-haves beside one merged AI summary, framed as decision support while the recruiter decides. Touch screens get 44 px targets, and every route was captured with accessibility (axe), overflow and touch-target checks.
I adopted passwordless managed identity and HMAC-signed calls between the web tier and backend.
For Azure Marketplace readiness, I packaged it as a single-organization Managed Application with signed releases and a private-network option behind Front Door WAF, then fixed 12 install findings until test installs of public and private plans passed.