/A minimal, responsive and feature-rich Jekyll theme for technical writing. 2026-04-13T00:55:49-04:00 Cameron Webb / Jekyll © 2026 Cameron Webb /assets/img/favicons/favicon.ico /assets/img/favicons/favicon-96x96.png 2026-04-11T23:50:47-04:00 2026-04-11T23:50:47-04:00 /posts/multi-agent-march-madness-2026-the-post-mortem/ Cameron Webb

48 of 63 games correct. 1,560 ESPN points (81.25%). The multi-agent pipeline called the champion, the runner-up, and 3 of 4 Final Four teams — including Michigan over UConn in the title game. It wasn’t perfect. The bracket also called Akron over Texas Tech as its highest-conviction upset and got blown out by 20. And one buzzer-beater 3 from Alvaro Folgueiras wrecked the entire South region. S...

2026-03-18T06:53:42-04:00 2026-03-20T00:20:36-04:00 /posts/multi-agent-march-madness-strategy/ Cameron Webb

Overview Below is my multi-agent March Madness bracket strategy. Rather than throwing a single prompt at an LLM and hoping for the best, I split the problem across seven specialized agents — each focused on a different dimension of what makes teams win or lose in March — then ran them through a three-phase pipeline designed to eliminate bias and force rigor. Five researchers work in parallel, ...

2026-03-07T23:30:16-05:00 2026-03-19T01:53:47-04:00 /posts/maximizing-ollama-inference-speed-on-m4-mac-mini/ Cameron Webb

This doc was generated with the help of Claude for researching how to improve response times of running Qwen3.5:9B on my Mac mini m4 (24GB) using Ollama. Posting this as a guide and resource for referring back to in future The base M4 Mac mini with 24GB unified memory can generate tokens at 28–35 tokens/sec with 8B-class models using Q4_K_M quantization — fast enough for responsive interactiv...