I build LLM systems that go to production and stay there — agent architectures, the evaluation stacks that prove they work, and the guardrails that keep them honest.
“Fifty analysts were filing tickets to ask questions about their own data.”
“The model could see the page. It could not tell you which box was the submit button.”
“Compare every market to every other market and the small ones all look like errors.”
First author
I'm Varun Malhotra, a Data Scientist · Generative AI & LLM Engineer at Michelin in Pune, India. My work sits across three layers of the same problem: agent architecture — deciding what an LLM system is allowed to do and when it should stop and ask; evaluation — building the benchmarks that prove a system works before a user finds out it doesn't; and guardrails — the read-only boundaries and human-in-the-loop gates that keep an autonomous system honest in production.
On Talk to Data I own the agent architecture, LLMOps, security and frontend workstreams — I architect, build, and engineer the system end to end.