About
I'm Kislay Kumar. I'm a senior software engineer on the Data and AI team at Lyric, a supply chain technology company, and I'm based in Bengaluru.
Six years in engineering, the last three of them building with large language models. At Lyric I built a multi-agent system that automates a large part of our feature development workflow, and the retrieval-augmented generation framework our debugging models run on. Both are used daily across a 350-person company, which has meant less time on prototypes and more on the questions that decide whether an LLM system holds up once people depend on it: retrieval quality, cost, evaluation, and the failure modes that only appear at real usage.
Why this blog exists
There's a gap in writing about AI. On one side is content that reads easily but says nothing — analogies that collapse the moment you push on them, no equations, nothing you could implement. On the other side are the papers, and writing that assumes you've already read them.
The useful middle is writing that's genuinely pleasant to read and mathematically honest. Not math-free. Math-explained. You should finish a post able to say what a query, a key, and a value actually are, why there's a √d_k in the denominator, and what would break without it.
That's the standard I'm writing to. A few rules I hold myself to:
- Never fake an explanation. If I don't understand why something works, the post says so.
- Every equation gets three things — the formula, what each symbol means in plain words, and what would go wrong if you removed that term. The third is the one everyone skips, and it's where the understanding actually lives.
- Intuition before formalism. Feel the problem before you see the solution.
- Cite the primary source and link it. These are real papers by real people.
I'm not a model trainer and I don't pretend to be. I'm an engineer who ships these systems and wants to understand them properly — which, as it happens, is exactly who I'm writing for.
Elsewhere
If a post made something click, or if I got something wrong, I genuinely want to hear it. The second one especially.