I design the platforms engineers and AI agents build on.

Turns any markdown plan into a page where I select text and tag it: wrong, refine, unclear.

• I submit a round, and the agent applies every note and records what it did for each one.
• The next round shows changes by section.
• If no agent is listening, a headless applier picks the round up. An approval to implement never runs unattended.

• Python standard library
• Vanilla JS
• Mermaid

Before a plan gets built, several different models attack it looking for holes.

• Any critique that quotes text missing from the document is thrown out.
• Surviving critiques get a defense round before anything is dismissed.
• It runs on local models, so the document never leaves my machines.

• Python
• Ollama

A memory for my agents, built on my notes.

• An Obsidian vault is indexed for search and kept in sync by a file watcher.
• It has a REST API, a CLI and a Claude Code integration.
• It runs in Docker at home.

• Node
• Express
• FlexSearch
• Docker

My kid types an idea and a local model turns it into a Minecraft add-on that installs in the game.

• The model writes a structured spec, and its scripts are type-checked against the game’s official typings before anything is built.
• A failed check goes back to the model with the reason, up to three attempts.
• Textures are generated in code, so nothing copyrighted ships.

• Node
• Express
• zod
• TypeScript
• Ollama

A chat my kids can use with local models, with safety in layers and a parent who can see everything.

• A keyword layer runs before any model, then a model screens each message in and out.
• The child’s age shapes both the screening and the answers.
• Blocked turns and transcripts go to a parent page, and push alerts carry no chat text.

• TypeScript
• Express
• SQLite
• Ollama
• ntfy