Build series

VERA

Verified · Explainable · Responsible Assessment

An honest account of building an AI hiring assessment platform from scratch, the architecture decisions, the things that broke, the research behind the fairness philosophy, and why this product exists at all. Written as it gets built.

All posts

01
Why I Built VERA: The 2AM Rejection That Started Everything
I got rejected by HireVue at 2AM. The system's own data said I was exceptional. No explanation was given. Here is what I found when I went looking for one and what I decided to build instead.
02
How VERA's Scoring Pipeline Works: Two Passes, Four Parallel Calls, and Why LLMs Cannot Do Maths
The naive approach was tempting. I tried it. It failed miserably. Here is why a single-prompt architecture breaks down under production constraints, and how we engineered a multi-pass, parallelised, schema-enforced inference pipeline that delivers compliant, assessment-ready frameworks in under 10 seconds.
03
The 429 Wall: Hidden Regional Quotas, a Data Residency Tradeoff, and the 16-Second UX Ceiling
VERA's production environment ground to a halt mid-interview. The culprit was a regional quota we couldn't even see. Here is how we pivoted our infrastructure, discovered an unexpected model capability finding, and engineered a 16-second UX ceiling to keep candidates from getting stranded.