Himanshu

I’m Himanshu, a Principal AI Specialist based in Bengaluru. I lead the AI business unit at MResult, where I design and ship GenAI and agentic systems for pharma and life sciences. The work sits in the awkward, interesting place where a language model meets a regulated process: document intelligence over pharmacopoeial and clinical text, supply-chain forecasting agents, pharmacovigilance assessment, and the validation scaffolding that has to exist before any of it is allowed near a real decision.

I came to this from mechanical engineering. I spent my first years writing reliability and fault-prediction software for oilfield equipment, taught myself data science on the side, and never really stopped teaching myself things. Most of what I know about deep learning, probability, and GPU systems came from books and my own experiments rather than a classroom, which is probably why I’m stubborn about building from first principles - pure Python, typed contracts, no orchestration framework standing between me and the thing I’m trying to understand.

The through-line in my independent work is reliability: how do you know a GenAI system is actually right, and how do you prove it to someone whose job depends on the answer? That question has turned into a series of small empirical studies on retrieval failure, on graph versus vector RAG, on Bayesian reward modeling where the finding matters more than the framing, including when the finding is that the clever approach didn’t win. Those write-ups are under Professional, and every one of them links to the repository that produced it.

Away from the screen, I read. I read constantly - probability, machine learning, and whatever else earns a place on the shelf and I keep a running list of it here. I have an unreasonable conviction that everything is physics, a long-standing argument with anyone who tells me “technology is not the problem”.