I build and debug the infrastructure that agents run on. Most of my recent work is on agent evaluation harnesses — the layer between a model and a task where trials fail for reasons that have nothing to do with the model, and where a bad harness quietly turns an infrastructure timeout into an "agent failure" in your results table.

Before that: backend engineering on distributed systems, and a degree in financial mathematics. Both turn out to be the right training for this. Evals are a distributed systems problem wearing an ML costume — sandboxes, retries, streaming, idempotency, partial failure — and deciding what a noisy score actually tells you is a statistics problem.

• Agent evaluation harnesses — open patches to Harbor covering sandbox lifecycle under custom task ENTRYPOINTs, gRPC exec-stream exhaustion on long runs, retaining failed attempts for post-mortem instead of deleting them on retry, and separating genuine model safety refusals from provider-side content-filter errors
• Training infrastructure — Modal serverless GPU runners and Weights & Biases sweeps wired across my RL repos, so an experiment is one command and a permalink
• Optimizer comparison for tool-using agents — DSPy optimizers benchmarked head-to-head on tool-selection tasks

• world-models — chess and Go distilled from Stockfish and KataGo on a single GPU: 2,301 Elo chess (95% CI [2,190, 2,601]) and ≥2,366 9×9 Go-Elo (GnuGo-anchored, not the AlphaGo-paper scale), at roughly three to four orders of magnitude less compute than AlphaZero. Includes the ablations and the negative results — MuZero from scratch caps at ~700–900 Elo at this budget, and ungated self-play from a distilled teacher regresses ~370 Elo.
• rl-playbook — rlplaybook.com, a visual timeline of deep RL's landmark papers from DQN forward
• rl / tabular-rl — RL agents with nothing abstracted away, from tabular methods up through PPO
• cassandra-playground — Cassandra-style anti-entropy repair in Go: gossip protocol with Merkle trees
• Project-Nash — Nash equilibria, Lemke–Howson, minimax, simplex; the game theory I keep reaching for in multi-agent settings

Distributed systems and infrastructure — gossip protocols with Merkle trees, Temporal.io workflows, Terraform on AWS. Quantitative finance — portfolio optimization, stochastic programming, derivative pricing. Python, Go, Rust, TypeScript, Java.

Seattle. Longer-form writing at shehabyasser.com.