In the open
Products currently live.
AgentWrangler
A local dashboard for what your Claude Code sessions cost. It reads your session transcripts on your own machine and shows where the tokens go and whether a change helped. No account, no telemetry. Built offline (3,686 commits across 57 active days) and released on GitHub and npm as a curated export of that private repo.
ReadySetBind
Turns an insurance quote PDF into a ready-to-sign bind package — the AI reads the fields, a human verifies every one, then it routes to the underwriter. A live pilot.
TariffRefunded
Tariff-refund recovery for small importers. Protest windows close on a fixed legal schedule and the refund pool shrinks 8–10% a month, so the whole product races the clock. Free analysis; broker partnership in progress.
Latest in the build journal
The three most recent entries. See the full journal →
- OCT 12
CareerHub — a job search run like a sales pipeline
A local-only pipeline for my own job hunt: ingest postings, score them, draft outreach, and a human Approve on every send. The first three commits were the spec, the decision log, and a review pass, before any code. 1,352 commits across 15 active days by Oct 29.
- SEP 4
AgentWrangler goes public
First release on GitHub (Apache-2.0) and npm the same day. The public repo (a placeholder since Aug 20) now carries a curated export tree, not the private history, and every later update ships as a gated export PR.
- AUG 21
AgentWrangler — first commit, offline
A local dashboard for Claude Code token spend, built on a private repo that never leaves the machine. By Oct 21 that repo held 3,686 commits across 57 active days.
Recent writing
Field notes from building with AI agents. All posts →
AgentWrangler is out. GitHub shows 78 commits; the real number is 3,686
The local Claude Code token dashboard I teased in August is public on GitHub and npm. The public repo is a filtered copy of a private build, and nothing reaches it without passing scans for secrets and personal data.
The cheap model failed its own exam
CareerHub is the private pipeline I built to run my own job search. The most useful thing it did was fail a cheaper model on an exam it couldn't cheat on.
Let Claude conduct, let Codex build
Claude workers were draining my weekly allowance. I moved implementation to Codex and kept Claude planning and reviewing. Here's what the token logs showed, what broke, and how the skills have changed.