About

Swedish-American, dual citizen, raised in Boise. Madison now. San Jose in the fall.

education

B.S. Data Science, CS minor, University of Wisconsin-Madison, May 2026. M.S. Computer Science at Northeastern University’s Silicon Valley campus in San Jose, September 2026 to May 2028. I picked the campus over the ranking: the work I want to do happens within twenty miles of it.

what I work on

My main project is thaw, git for live LLM agent sessions, built in Rust, CUDA, and Python. It checkpoints, branches, diffs, and restores live inference state (weights, KV cache, prefix-hash table, scheduler). A running vLLM session forks in 0.88s median on an H100 against a cold boot of roughly 340s. Getting there required a double-buffered O_DIRECT DMA pipeline that overlaps disk reads with PCIe transfer, an 8-shard parallel CRC32C verifier proven to match a serial pass, and coalescing about 16,000 tiny per-block DMAs into one contiguous gather to remove a 60x snapshot bottleneck. 16 releases on PyPI as thaw-vllm, 388 tests in CI (155 Rust, 233 Python) that run with no CUDA toolchain.

That work went upstream: PR #44074 (pluggable sleep-mode backend abstraction, out of RFC #34303) merged into vLLM core in July 2026, follow-up #47243 merged the same day. It turned into a vLLM open-source fellowship, sponsored by Inferact: July is engine cold-start, August is model hot-swap, and the first deliverable was a measured H100 phase ledger of where vLLM boot time actually goes (it also surfaced an upstream bug, fixed in #47356).

research

The question I keep returning to is state. A transformer mid-generation is a multi-gigabyte live data structure, and almost everything downstream of it (agent frameworks, RL pipelines, evaluation harnesses) treats it as disposable: throw it away, re-feed text, hope the numbers come back the same. thaw is the engineering half of an argument that it should be a first-class artifact. The measurement half is a sole-author arXiv paper, “Re-feeding Is Not Replaying” (June 2026): on stock vLLM, the re-feed shortcut changes per-token credit estimates at decision tokens 14 to 28 points above a replica noise floor, the perturbation is consistent with mean-zero so averages mostly survive, threshold-based token selection does not, and vLLM’s batch-invariant kernels eliminate the whole effect bit-exactly. Total compute was under $10 of rented A100 time, which I consider part of the result.

The plan from here: keep doing measurement work on inference systems, get the preprint through a workshop, and eventually a PhD in computer science. I want the kind of career where the deliverable is a number with a confidence interval and a repo attached.

founder

Matteson Systems LLC is the entity behind thaw, and also behind a separate product: an autonomous outreach system that audits local-business websites, runs Lighthouse Core Web Vitals in a real headless browser, does a Claude-vision pass, and writes the owner a personalized scorecard. 10,500+ businesses scored, 158 high-priority leads in one run, about 3 cents each. I applied to YC for the S26 batch; the application was rejected but placed in the top 10%, YC encouraged a reapply, and I am reapplying.

other

I make music in Ableton, play guitar, run, and ski when the snow is worth it. The site looks like Windows 98 on purpose; desktop computing peaked then and I see no reason to pretend otherwise.