AI tools

Tools I build for running AI-era teams.

Built on my own time and hardware. Each one is a piece of the system I teach: shared context, a delivery harness, spend visibility, and a report that measures what matters.

Live demo

Engineering effectiveness report

AI spend, forecast and harness practices for an engineering org, built for directors and for the engineers themselves, never for ranking people.

Fiscal-year spend forecast with an 80% range, budget variance by team and manager
At-bats: shipped experiments with a measurable hypothesis
Per-engineer savings: model right-sizing, prompt caching, rtk and context hygiene

Runs on a synthetic eight-team organization.

Open the live demo

Preparing release

knowledge-mcp

A local knowledge base for agents: remember decisions, business rules and findings once, and let every session search them.

Search, remember, update, forget and ingest tools over MCP
Hybrid retrieval: vector plus keyword, merged with reciprocal rank fusion
Local embeddings, so nothing leaves the machine

Python
MCP
SQLite
sqlite-vec

Notify me when it's released

In development

agent-harness

A clean-room, open-source delivery harness: phase-gated skills, specialist subagents, a worktree per ticket, and humans approving the parts that matter.

Triage, plan, review, execute, test, PR, deploy, close
Complexity gates decide when a person approves the plan
Agents never merge

Claude Code
Skills
Subagents
Git worktrees

Notify me when it's released

Preparing release

llm-usage-analytics

Turn LiteLLM gateway logs and agent session history into spend per team, model and month, and a picture of how people actually use agents.

Classifies sessions by type from the tools used and files changed
Imports Claude Code and opencode history
Feeds the effectiveness report

Python
LiteLLM
Postgres

Notify me when it's released

Want a walkthrough of how these fit together?

I show the whole system in talks and teach it hands-on in workshops.

Invite me to speak Book a workshop