I build things with LLMs and then break them until they work properly.

Right now I'm an AI Engineer at IFF, where I somehow ended up wearing every hat — architect, platform engineer, the "why is this costing us $40k/month" guy, and occasionally the person who deploys things at 2am.

Lately I've also been putting on the product hat for our enterprise LLM Gateway — talking to teams, figuring out what's actually painful, and shaping the thing into something people want to use, not just something that exists.

Enterprise LLM Gateway — the boring-but-critical infrastructure that sits between your team and the LLM providers. Governance, cost visibility, smart routing. The stuff nobody wants to build until the bill arrives. I'm owning this end-to-end: the architecture, the product decisions, the roadmap.

Autonomous agent orchestration — multi-agent systems that actually finish their tasks without a human babysitting them. MCP-style tool-calling, sandboxed remote execution, the works.

A platform where scientists just... run their ML jobs — talking to end users about their problems, finding better solutions for the people working on it, and facilitating a place for ML deployment. SageMaker, Kubernetes, Terraform, autoscaling. They click a button, models train. I lose sleep over the cost.

I don't have a "stack" — I have a problem, and then I find whatever solves it.

• Monday: building an LLM router that picks the cheapest model that's still good enough
• Tuesday: debugging why an agent decided to query Snowflake 47 times in a loop
• Wednesday: writing Terraform for the 900th time
• Thursday: talking to scientists about why their pipeline is slow and sketching a better architecture on a whiteboard
• Friday: asking myself why I didn't just become a surfing instructor

python · fastapi · kubernetes · terraform · aws (too much of it) · openai/anthropic apis · langchain · react · docker · go (learning) · postgres · grafana

Surfing · skiing · hiking · biking · reading papers about things I'll never implement · explaining to my family what I do for a living (still failing at this one)

If you're building something interesting with LLMs, agents, or ML infrastructure — or if you just want to argue about whether RAG is dead — I'm always down to chat.

mandipat@usc.edu · linkedin(something's never change, it's not up to date)

_{Currently obsessed with: making agents reliable enough that I trust them to run overnight without waking up to a $10k bill}