कामWork

Client systems, built to hold.

Two systems, two very different problems. Each one is told as a story: who it was for, what was at stake, and what I built.

Case १ · Mar 2026 - Present

IPO Allotment Engine

For Niche Technologies Pvt. Ltd. · SEBI-registered Category I RTA, via JRat's Studio
As Software Engineer (freelance)

प्रश्नThe problemMillions of IPO bids land when bidding closes. Under SEBI's T+3 rule they must be validated, reconciled and allotted overnight, with zero tolerance for error.
मार्गWhat I builtSplit a Go control plane from a vectorized Polars data plane on AWS Batch, and made the allotment lottery seeded and reproducible.
फळWhat it changed1.6M+ records reconciled per second (10M in 6.34 s), with database lock contention down 80%.

Case २ · Oct 2024 - Jan 2026

For Omara Technologies · Enterprise AI and document intelligence
As Founding Full-Stack AI Engineer

प्रश्नThe problemDocumentation, AI evaluation metrics and workspace permissions were scattered across products and tenants, with no single place to see how models were really performing.
मार्गWhat I builtBuilt one governance portal: Cognito RBAC across workspaces, a Ground Truth scoring engine, and the hub for DocuNexus knowledge graphs and the labelling platform.
फळWhat it changedEvery evaluation measured with field-level precision, recall and Cohen's kappa, and model edge cases fed straight back into training.

Case ३ · Aug 2026 - Present

For AgentDiff · My own product, open source and MIT licensed
As Founder & sole engineer

प्रश्नThe problemAgents fail quietly. A prompt tweak or model update makes the agent loop, waste tokens or cost more, and the test still passes because the final answer looks fine.
मार्गWhat I builtCompare two agent runs as aligned graphs, deterministically and locally, with no LLM judge, and turn drift, loops, cost and recovery into CI gates.
फळWhat it changedA pre-release on PyPI with a GitHub Action that blocks the pull request and names the exact step that regressed.