Portfolio

Brian Woods

Staff / Principal ML & Systems Engineer

UC Berkeley graduate: M.S., Information & Data Science, AI and machine learning concentration (2026)

I work on causal inference from messy observational data and the production systems that serve what I model: live trading infrastructure, drift monitoring, and published research with the negative results left in.

LinkedIn GitHub Blog

Latest study What is an H100 actually doing? Five open models on one rented H100: measure the card's real ceilings, predict each model against them, report the gap.

Open source

vLLMgithub.com/vllm-project/vllm

Correctness work across the engine, scheduler, and KV-cache paths. Found and fixed a prefix-cache bug in the V1 scheduler: when a session truncates tokens, a hash covering a discarded token stays in block_hashes, so a later request whose prefix matches the old tokens can be served KV computed from the new ones. The fix is one line in _update_request_as_session with a regression test that fails without it (#49377, #52806). Review contributor on a KV-cache capacity fix, where the review surfaced a decode-path livelock that became the fix's second commit (#52530, writeup), and on a stop-string leak traced through the detokenizer and reasoning parser (#52599).

H100 Roofline Studybrian.biz/h100

Roofline study of five open models on an H100: measure the card's real ceilings, predict each model's decode rate from its weights, report the gap. Qwen3-8B holds 81% of measured bandwidth. A concurrency sweep against one server turned out to be running at a 98% prefix cache hit rate, overstating throughput by 86%, filed upstream as vllm#52884.

Pydantic AIgithub.com/pydantic/pydantic-ai

Filed and fixed correctness issues in the agent framework's embeddings and AG-UI streaming surfaces: EmbeddingSettings.dimensions claimed Bedrock support that the underlying Titan models do not provide, and AGUIEventStream failed to emit agent-generated FileParts.

Agentic Workflowgithub.com/brianosaurus/agentic-workflow

A human-gated coding-agent pipeline: Claude builds, Codex audits in a read-only sandbox, and a human approves at every gate. Approvals are SHA-256-pinned to the artifact that was reviewed, and reviewer output is immutable, so an approval cannot be quietly reused for different work.

Live systems

Leeroy Chainkins/leeroy

The operations dashboard for my automated Solana trading systems: running bots and their positions, on-chain wallet balances, per-trade execution timing (build, submit, and on-chain fill), slot-level latency, and live logs. It monitors the infrastructure the Solana research below grew out of.

Machine learning research

Independent studies, each built on public or self-collected data with the full pipeline, tests, and negative results published.

What is faster block propagation worth on Ethereum?/optimum

A causal study of block propagation latency and validator revenue, built from 216,000 mainnet slots of public Xatu and MEV relay data. Identification comes from the protocol's 4-second attestation deadline (a sharp regression discontinuity: correct head votes fall from 95.9% to 67.4% across it), and the value of a faster transport is priced through three channels, with roughly 90% arriving via MEV timing rather than attestation rescue. Includes a 245-day analysis of the vendor's only real deployment, which finds no adoption-shaped change in propagation.

The same question, asked of Solana/optimum/solana

Quantifies how block completion timing relates to Solana validator earnings, using two days of on-chain data collected from my own Frankfurt node. Within-leader-window fixed effects separate timing from leader quality; leader fees turn out to be the larger channel, at +0.00372 SOL per 100 ms of earlier completion. A modeled 6x transport projects to roughly $794k per year for the largest operators. The confounds are stated up front: a single vantage point, skip risk, and simultaneity between timing and demand.

What is an hour of audio data worth?/david.ai

A pre-registered ablation study of ASR training data value: how word error rate responds to quantity, label noise, acoustic quality, and speaker diversity, using LibriSpeech and Whisper-small across 35 planned fine-tuning runs and nine pre-registered hypotheses. The goal is an exchange rate between noisy and clean training hours. The experimental pipeline is complete and verified by 213 tests. The fine-tunes await GPU compute, so the page reports no results yet.

Are agent benchmark scores stable enough to trust?/deccan

Tests whether agent benchmark comparisons survive perturbations that should not matter: prompt paraphrase, tool renaming, distractor tools, instance resampling, temperature, and task order, across 156 episodes on three procedurally generated task families. All six axes shifted scores, and detecting a 3-point gap between models turns out to require hundreds to thousands of instances. The conclusion is a cost estimate: a comparison you can defend costs one to two orders of magnitude more than these are usually given.