Sameer Ankalgi
Agentic AI systems & GPU inference · Microsoft, Zürich
I build tooling that steers coding agents — session control, evals, token telemetry — and make LLM inference work on H100/A100 in production. Currently in GitHub AI Engineering at Microsoft.
Before AI infra: 12 years of distributed systems — including a 2 TB / 600M-record MongoDB migration for Daimler, zero downtime, 4–5K writes/sec sustained through the cutover. The CV has the whole run.
- 600M+ records migrated, zero downtime
- ~4–5K writes/sec sustained through cutover
- 1,054 attack cases in the pi-bench suite
- 60–70% Copilot token spend cut by octerse
Building
- wingman Steer live GitHub Copilot CLI sessions from a phone — Go daemon, end-to-end encrypted relay on Fly.io, SwiftUI iOS app. Protocol spec and ADRs in the repo.
- pi-bench PINT grades detectors; AgentDojo grades models. pi-bench grades the composed defense stack you actually deploy — 1,054 attack cases, every row reproducible in one command.
- octerse A token-saver for GitHub Copilot: context engineering plus spend telemetry, cutting 60–70% of output tokens under usage-based billing.
- All projects
Open problems
Six unsolved problems from production inference and agent work — the notebook I operate from. Among them: why served p99 and benchmark throughput refuse to converge, and admission control when the KV cache is the resource that runs out first.
Production systems
- AlphaFold 2 on Azure H100 HPC — profiled the protein-folding pipeline end-to-end for a pharma customer and tuned throughput until large-scale structure prediction was viable.
- DeepSeek on multi-GPU A100 — served with vLLM and tensor parallelism, tuned for production throughput and latency.
- Daimler connected cars — the 2 TB migration ran live under a production write load. Nobody noticed, which was the point.
Writing
- Everything that tried to break my chess server Part 2 of the chessm8 story: the races I caught by attacking my own system, a board that lied to players, and launch-night bots hunting for credentials.
- Two players, one clock, zero trust: building a distributed chess server on the BEAM Part 1 of the chessm8 story: why one Elixir process per game, server-owned clocks, exactly-once ratings, snapshot resurrection, and one deliberate machine.
- Your prompt injection defenses may not compose I benchmarked stacked prompt injection defenses across four open-weight models. Two defenses that each work fine alone can block 100% of benign traffic when you chain them.
- All writing