hey, I'm pritam.
I care about the machinery underneath modern ML — the serving stacks, schedulers, kernels and memory systems between a model and a request. Frameworks are easy to use and hard to see into; most of what I do is an attempt to see into them.
technical interests
- Inference engineering — serving systems, scheduling, KV-cache management, continuous batching.
- ML systems — how models, runtimes and hardware interact, and why the fast path is fast.
- GPU performance — memory hierarchies, kernel behaviour, and the arithmetic of making hardware do what the roofline says it can.
- Post-training for creative domains — teaching models to be interesting, not just correct.
current focus
Right now I'm working through LLM inference and serving systems: reading implementations, running my own, and rebuilding simplified versions to test whether I actually understand a component or just recognize it. The honest, up-to-date list lives on the homepage.
how i think
First principles over familiarity: break a system into its fundamental components and reason up from there, rather than memorizing how a tool wants to be used. When something feels like a black box, I build a small version — an autograd engine, a minimal serving loop, a tiny VM — until the design choices become obvious. Then I experiment until behaviour is explained, not just observed.