I build AI products that hold up in production 
at startups that needed them working yesterday.Worked at
I split my time between why anything works at all and shipping things people use</>.
SalesDuo · Internship
Building features across a multi-service architecture: a multi-tenant SaaS core (Express, TypeScript, Sequelize) and an LLM-powered listings and brand-store optimizer (Python) behind a shared nginx gateway. Shipped end to end on AWS (ECS, ALB, RDS Postgres, Redis) with GitHub Actions CI/CD.
Softworker AI · Internship
Worked on the core of an agent runtime rather than the chat sitting on top of it: memory and context management, ReAct-style reasoning loops, and the orchestration that turns one task into runs, steps, tool calls, approvals and retries. Explored vector and graph stores (Qdrant, pgvector, Memgraph) and agent memory systems (Mem0, Letta, Graphiti) to work out how long-running work holds its state.
MarkupX Brands · Full-time
Shipped an enterprise GenAI platform with 12+ AI features like Virtual Try-On and Face Swap. Cut generation latency 40% with GPU processing, dropped re-renders 70% with Zustand, scaled to 100M+ records on MongoDB and Cloudflare D1, and integrated 40+ APIs including FAL.ai and Razorpay.
Selected
Most of engineering is deciding what you are willing to lose.
AI finance controller · Razorpay AI Buildathon
Matches invoices to bank payments, and is honest about the ones it cannot. The model never does the arithmetic and never has the last word: it recommends, and hard-coded rules decide what that recommendation is allowed to become. Zero wrong approvals across the held-out set.
Agentic personalisation pipeline
Four agents run in order inside an isolated microVM: check consent, pull the facts out, find matches, then one last agent checks the answer against the real data before it ships. Old facts fade with time, and ones that contradict each other get caught.
Adaptive rate limiter for Node
Mix any algorithm with any store. When the event loop starts lagging it drops from Redis to memory on its own, so a busy server sheds load instead of dying.
Writing
Writing it down is how I find out whether I actually understood it.
Now playing
Divyansh FM
Calm Ambient Dreamscape
morgan-ambient
what’s on while i build
