Building RAG Systems That Actually Work in Production
Vector search is the easy part. The hard lessons are in chunking strategy, context window management, and why naive similarity search fails medical data.
Tech Lead · Full-Stack Architect
Lead engineering for the company's flagship B2B product, joining at a point when the in-house engineering function was close to non-existent. Built the team and its delivery practice alongside shipping, and act as the technical voice in client-facing discussions across the company's project portfolio. Engaged through October 2026, when v1 is planned.
What I did
Built here
Engaged independently to take the company's catalogue onto AWS Marketplace. Designed and built the full-stack listing and SaaS billing integration for seven products, including automated provisioning that stands up a customer's environment on purchase with no manual setup.
What I did
Built here
Lead end-to-end architecture and delivered production-grade systems serving 50K+ daily operations. Mentored 4 junior developers with 10+ code reviews per week, established coding standards, and spearheaded sprint plannings with 100% on-time delivery. Orchestrated monolith-to-microservices transitions, improving system maintainability and performance.
What I did
Built here
Full-stack development across multiple client projects. Built document automation, OCR systems, WhatsApp workflow bots, and community platforms.
What I did
Built here
Freelance Developer
Independent projects helping businesses streamline operations and improve software usability.
What I did
Built here
Nobody commissioned these. They run on infrastructure I pay for and administer, which is the only reason they are worth showing.
Backend Engineering
Frontend & Mobile
Data & Search
AI/ML Integration
DevOps & Tools
Leadership & Architecture
Friend Affinity Finder — social graph-based friend recommendation system
Devil's Advocate format, Jalpaiguri Government Engineering College (2018)
International Institute of Information Technology, Bangalore
Jalpaiguri Government Engineering College (MAKAUT)
Budge Budge Institute of Technology (WBSTCE)
Vector search is the easy part. The hard lessons are in chunking strategy, context window management, and why naive similarity search fails medical data.
We achieved 5× throughput improvement — but not by splitting services randomly. Here's the non-obvious decisions that made the difference.
Course: boot.dev — Learn Go
Go's const is compile-time only — and how that differs from Java, C++, Rust and JavaScript.
Incident: Portfolio performance pass — measure, then fix
My site scored 23 on mobile. I had three suspects and all three were innocent. The biggest cost was a single prop that downloaded a 1.5 MB file from someone else's GitHub on every page load.
Incident: Mini-Me voice stack — build, measure, replace
A GPU voice stack that worked, cost ₹520–620/month, and was never switched on. Five reasons it sounded robotic, and only one of them was the model.