James Heath

Gauge tracks your workouts, monitors recovery signals, and adapts your program in real time. The AI coach sees your PRs, readiness, and history — then adjusts volume, intensity, and exercise selection so every session fits how you're showing up that day.

One of the most amazing things about toddlers is how much they can do and manage on their own. Routines help immensely. While his iPad gathers dust between plane flights, I thought I'd put it to use: a pictorial routine calendar so that he can see and interact with his day, laid out in friendly cards — wake up, play, breakfast, get dressed — with educational mini-games mixed in.

Lodis proved that portable AI memory is useful, but the bespoke graph and cloud layer did not beat simpler structures like committed docs, ADRs, session notes, and lightweight memory conventions enough to justify the upkeep. The lesson was not that memory does not matter — it matters a lot — but that the durable source of truth should be boring, inspectable, and easy for both humans and agents to maintain.

An orchestrator that ended up too close to a clone of Claude Code itself. It might be worth repurposing into a DevOps/bug fixing tool, but my vision of going from me writing a PRD to Agent Forge writing finished software didn't work reliably. It also became expensive to run, consuming a ton of tokens via API vs. my (presumably subsidized) Claude subscription. On the shelf for now, but it earned solid learnings about structuring data and workflows for autonomous agentic work.

Everyone wants a personal assistant, right?! This was my first playground. The further I got with it the more I realized:

1. I was recreating UI surfaces and primitives that already existed elsewhere in better tools
2. It was token and time-intensive to build and maintain
3. I needed to build features out more slowly and deliberately to get them to really work well with an LLM under the hood.

My first real build: a bot that would grab hard-to-get reservations the instant they were released and snipe cancellations as they opened up. It turned into a cat-and-mouse game with bot detection and never got reliable enough to trust. The punchline came later — a one-off scheduled Cowork task driving an authenticated Chrome session actually does the job, as long as you're diligent about setting it up each time. Rez Sniper was trying to automate away exactly that diligence; it just couldn't clear the reliability bar. Killed it, but it's the project that taught me to build.

A menu-bar dashboard that puts your Claude and Codex usage in one place — session and weekly limits across both, when each one resets, your projected monthly bill, and what that same usage would've cost at API list prices. Forked from Claude God and bent toward the question I actually had every day: how much have I got left before I hit the wall? The name's a play on "token maxxing" — run hot, just don't run out.

What began as a simple log grew a watchlist, a recommender, and viewing stats nobody asked for. The real exercise was velocity — how complete a build I could get out of one tight iteration loop in a single sitting, empty repo to live URL. Complete enough that it stuck.