Work

vLLM (upstream + fellowship)

Two PRs merged into vLLM core in July 2026: #44074, a pluggable sleep-mode backend abstraction out of RFC #34303, with review engagement from NVIDIA Dynamo and Alibaba Cloud engineers, and follow-up #47243 (communicator-agnostic capability flags), merged the same day. A third, #47356, fixes a bug I found while measuring: the documented fast-boot flag silently invalidates the torch.compile cache, costing +21-27s to save ~2s. That work became a vLLM open-source fellowship, sponsored by Inferact: engine cold-start (July), model hot-swap (August), starting from a measured H100 phase ledger of where boot time actually goes.

thaw

thaw is git for live LLM agent sessions: checkpoints, branches, diffs, and restores live inference state (weights, KV cache, prefix-hash table, scheduler). A running vLLM or SGLang session forks in 0.88s median instead of a ~340s cold boot on an H100. The core is a double-buffered O_DIRECT DMA pipeline overlapping disk reads with PCIe transfer, plus an 8-shard parallel CRC32C verifier proven equal to a serial pass. Open source, 16 releases on PyPI as thaw-vllm, Apache-2.0. Full engineering writeup, receipts, and the honest negative result on its own page.

Read the thaw writeup / thaw.sh / repo

Matteson Systems LLC

An autonomous outreach system for local businesses. Scans a region from OpenStreetMap (10,500+ businesses scored in one run, 158 high-priority leads surfaced), screenshots each site in a real headless browser, runs Lighthouse Core Web Vitals, and does a Claude-vision pass to name the single most damaging fixable problem. Writes the owner a personalized email and a plain-English scorecard. Append-only event-sourced Postgres CRM with a threaded Gmail drip on Vercel cron. About three cents per business. Built solo.

Read the writeup / mattesonsystems.com

Research Cyberinfrastructure, UW-Madison DoIT

AI Workflows Research Assistant, January to April 2026. LLM evaluation and cost-tracking framework on AWS Bedrock: 9 models, 3 ensemble strategies, 282 questions. Ensemble majority voting scored 0.840, beat every individual model. Pareto analysis: Llama-4 Maverick at 98% of top accuracy at a fraction of cost and latency. Presented at UW-Madison ML+X Forum.

Selected projects

Sentinel. Distributed message queue in Go, Kafka-inspired, built from scratch: custom LSM-tree storage engine, skip-list memtables at 3.9M reads/s, SSTables with CRC32, write-ahead log, leveled compaction, gRPC wire protocol with topic and partition consumer groups. Raft consensus hand-rolled: leader election, log replication, split-brain prevention, tested against a deterministic in-memory network simulator. 45 tests. Read the writeup / repo

Madison Metro ML. Live ML that corrects the transit API’s ETAs. 47-feature XGBoost model plus Mondrian conformal prediction, calibrated 90% coverage interval stratified by route, day-type, and horizon. Retrains nightly through GitHub Actions behind a hard deploy gate (at least 2s MAE gain required). React, DeckGL, and MapLibre frontend renders 200+ live vehicles at 60fps. Read the writeup / madisonbuseta.com

Lattice. Reactive framework in Rust with a PyO3 binding and a Cranelift JIT, measured at 10.6x Streamlit on the same workload. repo

brain2text. BCI decoder for a Kaggle task: 5-layer GRU with a CTC decoder, 40% WER. repo

gitstare. Git-history TUI written in Rust. repo

LockBox. Password manager on AES-256-GCM. repo

Coursework

Distributed-systems visualizations of Cassandra, Kafka, HDFS, and HBase (CS544); 1.2M-tweet sentiment pipeline on the UW HTCondor grid (STAT405); causal-inference propensity-score matching in R (STAT479).