~ / projects
The work.
No demos. No résumé padding. Every project here is load-bearing — it runs daily, on owned hardware, and something breaks if it stops.
Hermes Agent Platform
A persistent multi-agent system: cognitive memory, scheduled autonomy running unattended for hours, conference rooms where agents debate design, voice in and out. The seventh full rebuild of the agent architecture — each version existed because the previous one hit a wall that couldn't be patched.
Case study covers: why seven rebuilds, the memory architecture, what unattended autonomy actually requires.
Fish Bite
A fishing forecast engine that models instead of guesses: solunar timing, weather fronts, thermal optima per species, water gauges — fused into a score that gets regression-pinned against real logged catches. When the model said "midday is dead," the catch log proved otherwise, and the model got rebuilt to listen.
Case study covers: physics-informed scoring, calibration against ground truth, shipping as a daily-use product.
Homelab Infrastructure
The backbone: one bare-metal node running stable services 24/7 — third generation of a lab that started as a used-workstation experiment, grew into a hypervisor with GPU passthrough, then consolidated when the workload proved the big iron was idle. Specs published at class level — enough to learn from, not enough to map.
Case study covers: three generations, GPU right-sizing, why every service self-hosts.
// retired but not forgotten
OpenClaw — the agent platform behind this site's previous life — lives on as a case study in Building Self-Improving Agents. Its six iterations are the reason Hermes exists.
read the postmortem →