I lead engineering on three production systems at RVM CAD in India: an AI hiring platform for skilled-trade workers, a job portal, and the operations platform that runs a four-branch training institute. I write much of the code myself, and the security, data and AI-access layers are where I do my best work. Before software I spent five years maintaining Airbus A320s for Air India, where a passing check has to mean exactly what it says.
Two connectors that let an AI assistant answer operational questions from live systems — without a database credential, without write access, and without ever returning a half-right number.
11 read-only tools on the job portal connector
Recording every prompt and completion on an AI hiring platform, so bad answers can be investigated — encrypted at rest, erasable per worker, and built so raw text cannot reach the recorder by accident.
6 AI call sites traced
A résumé generator for workers who have never had one. The sheet is printed and handed to a supervisor, so "it rendered" is not good enough — it has to fit one page, use the right fonts and never invent a fact.
56 / 56 fixture sheets on exactly one A4 page
Invited teammates on an employer account saw zero jobs, zero credits and zero candidates. The fix re-scoped fifteen business tables to the organisation — with no new column, no backfill and no change to the API.
15 business tables moved to organisation tenancy
Read-only connectors that let an AI assistant query production data through a scoped identity, with every call checked, bounded and logged.
Privacy boundaries, encrypted traces and deterministic checks around model calls, so a bad output degrades the result instead of corrupting the record.
Authentication and authorisation, multi-tenancy, payments, migrations and CI — on live systems where a deploy cannot wait for a maintenance window.
Tests and gates that can actually fail. If I cannot say what a red result would have required, I do not count the green one.
In progress · October 2026
Building a human-reviewed evaluation set from production, then shadow-running an open-weights model beside Claude Haiku on an NVIDIA DGX Spark pilot. Each task moves off the frontier model only after it passes its own gates on that set — schema validity, field agreement, Hinglish quality, latency — and Haiku stays as the automatic fallback.
Role families I fit, the scope of my current work, and how to reach me.
Short, scoped engagements on AI access, AI reliability and production audits, and training for engineers.