projects

Selected recent work. Full list at github.com/jconnolly.

AI / LLM

local-llm-pi5 · SOTA local LLM at home, two-track investigation: Raspberry Pi 5 16GB + Hailo-10H AI HAT+ 2 vs. a spare M2 MacBook Air. Verdict, with measured numbers: if you have spare Apple Silicon on the LAN, use it. qwen3:14b at ~10 tok/s on Maral beats qwen3:8b at ~2 tok/s on the Pi, for $0. Hailo-10H is good at compiled HEFs (vision, Whisper) but not at agentic LLMs — 2B param ceiling, 2k context, broken tools payload in hailo-ollama. Notes on tool-use parsing (qwen2.5-coder emits bare JSON, qwen3 emits proper tool_calls), memory guardrails so Ollama can’t thrash SD-card swap, and the Claude Code wiring.

spotify-cagebreak · Weekly “algorithm cage-break” music discovery. Dumps your Spotify taste profile, asks a local Ollama model for opinionated picks adjacent to your taste but not in your library and not recommended before, then builds a dated playlist and a markdown digest. No recommender-algorithm feedback loop — you get LLM taste, not Spotify’s.

Hardware / OpenWRT

google-wifi-suzyq-console-macos · Serial console + OpenWRT flashing for Google WiFi (Gale, AC-1304) from a Mac via a $7 SuzyQ/SuzyQable USB-C debug adapter. No Chromebook, no Linux laptop, no kext. Companion to the unbricking-six-google-wifi-pucks post.

Other recent

constradoak-vacation-hunter · 2027 Presidents Day family vacation hunter — destination research, airfare playbook, automated fare-watch monitoring notes.

Earlier (and ancient)

Grim et al., “Reducing Prejudice” (MIT Press, Artificial Life IX, 2004) · Co-author on a spatialized game-theoretic model of the contact hypothesis — agent-based simulation of prejudice reduction, done at SUNY Stony Brook under Patrick Grim. Full cite + artifacts on the about page.