Projects

Tools and experiments: the problems, implementation decisions and limits behind the code.

• Ban Comments: a configurable ESLint rule for comments

  How eslint-plugin-ban-comments enforces a team convention, preserves deliberate exceptions and removes comments without pretending to refactor the code.

• Chest: evidence-based completion checks for AI agents

  How Chest uses local criteria, validation reports and explicit verdicts to make AI agent work inspectable—and where its fallback checks need care.

• Crew: why I stopped building an AI development platform

  A Crew case study: four months of implementation, a small agent benchmark with higher token and time costs, and the decision to stop before an external pilot.

• Fake-End: mock APIs with YAML and TypeScript

  Build and test frontend states before the backend is ready using Fake-End's file-based routes, request interpolation, guards and local API fixtures.

• Logbook: durable task state for AI agent handoffs

  How Logbook preserves tasks, context and session ownership across AI agent handoffs, with readable JSONL records and lifecycle hooks.

• Ohtools: inspect tool contracts before agents execute

  How Ohtools turns a shared tool registry into an inspectable hierarchy with schemas, next-step context, CLI access and MCP integration.

• Plot: layered specifications and exact-target approvals

  How Plot connects concept, umbrella and feature specifications through inheritance and approval gates—and what its scaffold currently exposes.

• Sounder: explainable Fibonacci estimates for AI tasks

  Follow a verified Sounder estimate from task text to weighted signals and a Fibonacci bucket, then inspect feedback and model-inventory limits.