Selected work
I build AI tools, and the layer that gets them adopted.
My work is AI enablement: the tools that make a model useful to a real team, and the systems that keep it honest once it is running. Below are the ones I can show. The enablement tools are public: installable, source-readable, tested. The systems I built and ran myself, and I kept the numbers.
A fair question from a hiring manager: I run a small practice, so am I about to leave? No. The practice is the workshop where I keep this work current and measured on real problems. A role is where it runs at the scale one person cannot reach, with a team and stakes bigger than a client list. The practice is how I stay sharp for the job, not an exit from it.
Enablement and adoption
Tools that help people actually use AI, shipped in public so you can check them yourself: install them, read the source, run the tests.
A decision tool that tells you whether to hand a task to AI, and hands you a stop-test instead of hype.
A no-code eval harness: write test cases in plain English, get a report card and a ship / fix-first / do-not-ship brief.
Point any MCP client at a folder of markdown SOPs and FAQs, and your team can search and cite its handbook. Two-minute, no-code install.
Copy-ready prompts for customer-success work, each with a five-element teardown you can actually trust.
Reliability, cost, and evaluation
Public tools that keep AI honest and affordable once it is running. Each ships its raw data and a one-command check that re-derives every number offline, so you can verify the headline yourself.
A cost auditor that prices real Claude traffic against the vendor rate table and reports, in dollars with a confidence interval, what prompt caching actually saved.
An eval harness that tells you whether a model swap is a real regression or just the model disagreeing with itself, with the statistics to back the call.
Post-mortem tooling that turns a Claude Code transcript into a redacted record and points at the exact turn where a run broke.
Systems I built and ran
Production I designed, ran, and measured myself, under real load, with the receipts kept and reproducible.
A job-search intelligence system: it scores thousands of postings and weeds out the fakes.
A set of rules and review agents that keep an AI coding assistant honest.
A reliability layer for LLM work: tier routing, human-gated loops, computed confidence, and claim ledgers, so an agent knows where it might be wrong.
An MCP server that sells competitor data to AI agents instead of people.
One script that turns a short spec into a whole catalog of finished products.
More is in progress: a semantic cache that publishes its own false-hit rate, and a docs-drift auditor that only proposes a fix it can re-verify. Each lands here the same way the tools above did, finished and independently verified, with numbers you can check. Nothing goes on this page without a real result behind it.
Want the version wired into your own business, with the price next to each line?
See the rate card