Seven years across Node.js, TypeScript and React — now focused on autonomous agents, applied ML, and the event-driven backends that hold them together.
I'm a senior software engineer with seven years building production web and mobile applications — and, increasingly, systems that build and run themselves.
Today I'm an SDE II at Brightstar Lottery, where I built an internal AI agent that fixes bugs straight from Jira — opening its own pull requests with a human in the loop — and shipped complete internationalization for the mobile app. Outside of work, I design autonomous-agent, multi-agent, and applied-ML systems on event-driven backends.
What I care about most is the unglamorous layer: queues that don't drop jobs, releases that don't lose data, and logs you can actually trace. That's the difference between a clever demo and a system that holds up in production.
Queues that don't drop jobs, workers that auto-scale, and back-pressure that degrades gracefully instead of falling over.
State-preserving deploys with checkpoints and backups, so an upgrade never loses data and a rollback is always one step away.
Request-scoped, traceable logs from day one — because you can't fix what you can't see, especially when agents act on their own.
A phase-based pipeline that orchestrates LLM (Claude) subprocesses to build software on its own — per-project FIFO queues, auto-scaling worker concurrency, request-scoped logging, and a state-preserving release workflow so upgrades never lose data.
Independent GPT- and Gemini-backed agents coordinated over a shared messaging layer — one for sales workflows, one for market analysis. Each is its own service, so models and prompts swap per task without touching the rest.
Selects message variants in real time with a Thompson-sampling multi-armed bandit — balancing exploration of new variants against exploitation of proven ones — fed by behavioral scoring. Event-driven on BullMQ / Redis over MongoDB.
Production pipelines for LLM content generation, multi-engine image background removal (RMBG-1.4 / BEN2 / FalAI), and automated voiceover. Heavy jobs run async on queue workers and survive restarts, with object storage on Cloudflare R2.
A GST / TDS billing application for Indian tax compliance: GSTR-1 JSON export to the government portal schema, Razorpay payments, and Form 26Q / NSDL FVU generation — each demanding exact conformance to statutory file formats.
An in-browser product-photo editor built on Konva.js with a layered effects system, batch processing on queue workers, and multi-image PDF export — complex interactive frontend on top of a real processing pipeline.