Hi, I'm Arun
I take ideas from AI research and make them hold up in production. I write about AI agents, memory and retrieval, and pretty much anything else in AI worth testing, with the code and data so you can check every number.
A GenAI project succeeds or fails on whether you can name the business metric it moved — and prove it. The architecture is the price of entry, not the win.
· 7 min read
15 of 19 escaped · 79% escape rate on the sample run
~40% lower cost · NDCG@10 held at or above 0.75
Case studies in progress
I'm writing up a few projects properly — outcome, stack, and what I'd do differently. For now, the writing is the clearest window into how I work.
Newsletter
Experiments, explainers and build logs when they publish. Choose technical posts only, or everything.
I came to research from production. My questions start as incidents: retrieval that misses the second hop, evals that stay green while the system gets worse, agents that succeed by accident. I study them with a stated hypothesis, a fixed setup and an interval on every number, and I publish the null results too.
Research interests · retrieval over structured knowledge · LLM evaluation methodology · agent reliability
More about me
