About
I learn by building the smallest thing that actually runs.
I'm Sudip — an AI engineer who's happiest taking something I half-understand from a paper and turning it into a tiny reproduction I can poke at. I care less about the leaderboard number and more about knowing exactly why a system behaves the way it does.
Day to day I work on inference optimization at the hardware level, new model architectures, and agent orchestration — kernels and memory bandwidth, papers that rethink the computation instead of just scaling it, and multi-step tool-using systems. This site is where I keep the receipts: what I tried, what broke, and what I'd do differently.
What I'm into right now
Hardware-aware inference
Kernels, memory bandwidth, and the tricks that make training and inference faster on the GPUs I actually have.
New algorithms & architectures
Chasing papers that change the computation itself, not just scale up the same one.
Agents & orchestration
Multi-step, tool-using systems — and figuring out where they quietly fall apart.
A short timeline
2024 — now
ML Engineer
Building and shipping LLM-powered systems; obsessing over latency and cost.
2022 — 2024
Software Engineer, ML
Data pipelines, training infra, and my first taste of "it works on the eval set."
earlier
Learning in public
Reimplementing papers for fun until it turned into a career.
Want to compare notes?
I like talking to people building in the same space.