nilsmatteson.com *** est. 2026 *** new: 2 PRs merged into vLLM core + a vLLM fellowship, sponsored by inferact (jul 2026) *** 0.88s session fork vs ~340s cold boot *** 14.3 GB/s weight restore *** sole-author preprint, "re-feeding is not replaying" (jun 2026) *** available for internships NOW *** fall 2026 onward *** zero js on this entire site. the marquee is HTML baby

writing - Notepad

Writing

Engineering notes, mostly about systems and applied ML. I try to name what broke, not only what worked.

Title Date
Echo and the Engineer A man with a chatbot says he has cracked cold fusion. I build these models for a living, and the part that unsettles me is not the man. A piece on sycophancy, RLHF, and why reality is not a term in the loss function.
Building a Real-Time Bus Prediction System for Madison Metro Live ML that corrects the transit API's ETAs with a 47-feature XGBoost model and Mondrian conformal prediction, retrained nightly behind a hard deploy gate.
Deploying RAG in AWS Bedrock: Benchmarking 9 LLMs on the WattBot Challenge Ensemble majority voting scored 0.840 and beat every individual model. The highest-citation model finished last. A serverless RAG pipeline on Bedrock with full cost tracking.
Building a Speculative Decoding Engine from Scratch Custom Triton kernels, tree-structured attention, four bugs, and an honest negative result: 0.66x baseline.