The name is a question I ask myself while editing.

I'm Shivam — I work on machine-learning systems, and I write here about LLM engineering and the unglamorous infrastructure underneath it. The day job is making inference fast and reliable; the evenings are for writing down the parts that were harder than the papers implied.

This site exists because the good stuff kept getting lost. A profiling session would turn up something genuinely surprising — a cache policy that behaved backwards under load, a retrieval metric that flattered a broken system — and it would end up in a Slack thread, then nowhere. Writing it down properly forces the second question: do I actually understand this, or did I just get it working? Usually the answer is uncomfortable, and the essay gets better.

Everything here is first-person and measured. If I claim a number, I ran it; if I'm guessing, I say so. I don't cover launches, I don't do hot takes, and I try very hard not to write anything that could have been a tweet.

Most of what I know I learned by being wrong in public, slowly.

Most of the writing is free. A few of the longer essays sit behind a one-time unlock — a single payment for that piece, set per essay, not a subscription and nothing recurring. It's less a growth strategy than a filter: paying a little usually means you intend to finish, which is the only readership metric I care about.

If something here is wrong, tell me. Corrections get a note at the bottom of the piece with your name on it, unless you'd rather they didn't.