Hi, I’m Keyur Ramoliya. I’m a software engineer and architect based in India.

Career history

1. 6ecbf87 merge haat
3. 33a6d8a haat Sr. Software Engineer Haat Delivery (2025–2026)
5. 0ab6671 merge casepoint
7. 3ffeaba casepoint System Architect Casepoint (2022–2025)
9. cbad523 merge datagrokr
11. f39f077 datagrokr Sr. Software Engineer DataGrokr (2021–2022)
13. dbbd8c4 merge casepoint
15. 3d8bc4f casepointv4 System Architect Casepoint (2021)
16. 8291a6f v3 Associate System Architect Casepoint (2020–2021)
17. 7117f63 v2 Sr. Software Engineer Casepoint (2018–2020)
18. 4182312 v1 Jr. Software Engineer Casepoint (2017–2018)
20. d3f2ba1 merge peacock
22. effae88 peacock Jr. Software Developer Peacock Technologies (2016–2017)
24. 0000000 Initial commit B.E. Information Technology (2016)

For the last few years, most of my attention has gone to AI, and to what it takes to run it for real: fine-tuning models, gating releases with honest evaluation, serving live traffic from self-hosted inference, and more recently building agentic systems with memory and code understanding. The question I think about most is how we make agents trustworthy as they start acting on our behalf. My answer so far: the rules belong in deterministic code, not in asking a model to judge itself. Close behind is a question I find endlessly fascinating: what is actually happening inside these models? To find out, I’m doing research in mechanistic interpretability, trying to reverse-engineer the circuits and features that models learn.

This blog is where I think out loud about all of it. As the line on my home page says, no one reads the blogs, but I keep writing anyway, because it’s how I find out what I really understand. When I’m not at a keyboard, you’ll usually find me with a book or a podcast.

If you’re working on similar problems, I’m always happy to talk. Find me on LinkedIn and GitHub.

All views and opinions expressed on this blog are my own.