Sameer Ankalgi

∣ ℎ𝑒𝑖𝑠𝑒𝑛𝑏𝑒𝑟𝑔 ⟩

AI infrastructure & distributed systems · Microsoft, Zürich

(this site has a shell — ↓ or press /)

I build AI infrastructure and distributed systems — the layer that decides whether a model ships or stalls. GPU inference, serving stacks, and the platforms underneath them, currently in GitHub AI Engineering at Microsoft.

Lately: profiling AlphaFold 2 on Azure H100 HPC clusters, and serving DeepSeek across multi-GPU A100 with vLLM and tensor parallelism. Before that, twelve-odd years of the same problem in other clothes — the CV has the whole run.

600M+ records migrated, zero downtime
~4–5K writes/sec sustained through cutover
12+ years across the stack
H100 · A100 vLLM · tensor parallel · CUDA

Building

pi-bench Grades whole prompt-injection defense stacks — ASR, false positives, latency, cost.
oss-model-playbook Running open-weight LLMs for real: capacity planning, gateway, security, TCO.
vector-engineering-for-agents The retrieval layer under an agent, at design-review depth.
All projects

Writing

The Coming Wave — a review Why scalable oversight has to outpace capability.
Why this site exists A small, slow, opinionated home on the internet.
All writing

Terminal

This site has a shell. Try help, nvidia-smi, or bench.

sameer@terminal:~$ █