personal workshop · local-first AI
Jake builds AI systems that live on hardware he owns.
$ currently building:
Agent orchestration, cognitive memory that actually remembers, a homelab that runs 24/7 — and shipped tools people use daily. This site is the workshop log: what was built, why it broke, what got learned.
selected work
Built to be used, not demoed.
Everything here runs daily on owned hardware. No proof-of-concepts, no portfolio filler — if it's listed, it's load-bearing.
Hermes Agent Platform
A persistent multi-agent system with cognitive memory, scheduled autonomy, and voice — the seventh iteration of an agent architecture rebuilt seven times because every limit demanded it.
Fish Bite
A fishing forecast engine with a physics-informed scoring core, thermal and pressure modulators, and regression-pinned calibration — built because the existing apps guessed instead of modeling.
Homelab Infrastructure
Hypervisor-free by choice: one bare-metal node running stable services 24/7 — the third generation of a lab that shrank itself on purpose. Keeps every other project alive through bad code, dead drives, and 3 AM experiments.
the architecture
Intelligence falls. Work gets done.
The whole shop runs as one cascade: frontier models reason and plan at the top, tasks fall through the agent pipeline, local GPUs handle execution, and a cognitive memory layer remembers everything worth keeping.
Frontier handles the thinking. Local handles the doing. Memory makes it compound. No part depends on a single vendor staying friendly.
read the architecture →- ▲ frontier reasoning — planning · verification
- ◆ orchestration — hermes pipeline · gates
- ● local inference — owned gpu · selective api
- ✦ cognitive memory — act-r · activation-weighted
interactive
This site has a terminal. It works.
A real command parser — navigate the site, read the manifesto, poke around. Click the window and start typing.
↑ history · try whoami ls projects neofetch cat manifesto.txt sudo hire-me
from the log
Recent writing.
The Catch Log Is the Reviewer
Two times the forecast model got caught lying, and the regression-pinned rebuild that made lying impossible again.
Agents That Build While You Sleep
Checkbox plans, fresh context per step, revert-on-failure: unattended builds that survive past hour three.
The Memory Layer: When an AI Actually Remembers
Activation dynamics, spreading activation, decay — and what changes when your assistant stops forgetting you.
Building Self-Improving Agents
Six iterations of an agent platform and the lesson each one beat into me: infrastructure beats intelligence.