Lab
I’m heads-down building and experimenting, especially on AI products where domain rules and checks sit around the model. Each project here is a receipt for a piece of that craft.
- Private beta
Ontrack
AI coaching for trail and ultra training.
ontrack.runWhat it proves
Coaching methodology lives in code as deterministic rules. The model helps with language and the exceptions. A plan that breaks the methodology does not ship.
- Open source · alpha
Writers Room
A writing studio that runs inside Claude Code and learns your voice.
GitHubWhat it proves
Voice becomes an auditable spec plus a local vault of facts and samples. Every edit I make turns into an evidence-backed rule for the next draft.
- In progress
Whisper Lakes
A cozy survival game about a boy lost in the wilderness. He catches fish with his fly rod and forages berries to stay alive. Inspired by Hatchet, which I read as a boy and just read to my son.
What it proves
Three.js on WebGPU can carry a real simulation. The rod and line cast with simulated physics. The 6 km by 6 km world runs real hydrology from alpine headwaters to the ocean, so trout and other species hold in the water where they would live. Each species has its own behaviour and micro regions.
- In progress
Fruit Shift
A real fruit-fly connectome works a five-minute shift, replayed live in 3D.
Open the viewerWhat it proves
A 25.6M-synapse research dataset, shaped into something you can watch in a browser.
- Open source
ChoreBank
A self-hosted chore and rewards board for the whole household.
GitHubWhat it proves
Small, opinionated, and finished. Next.js, Postgres, and Docker, MIT licensed.
- Live
Evil Chicken
A four-chicken 3D brawler, named after a playground game my son made up in kindergarten. He and his friends still play it at school.
Play itWhat it proves
AI can build a playable 3D game, the 3D models included. A small utility AI runs the opponents. Play with a paired Xbox controller for best experience. Mobile support is rough.
-
More coming soon.
How I think about harnesses
A model is good at language and bad at knowing what’s allowed. So the domain goes in charge. The rules that matter live in code, the model works inside them, and deterministic checks decide what ships. It takes longer to build the first time. It’s much faster to trust.
Building something where this pattern matters? Email is the fastest path. dhalarewich@gmail.com
Model draft 1 of 3
Tuesday: add a 24 km long run.
- Weekly volume rises 10% or less +22% fails
- No two hard days in a row OK passes
- Long run stays under a third of the week 38% fails
Blocked by 2 rules. The draft goes back to the model.