What I built for my employer, what I’ve built in code, and spec builds scored against answer keys. Client names are withheld; every figure comes from that build’s own runs.
The CEO’s email assistant and three agents, running at my current employer, and an archive assistant colleagues use in their own Claude.
Reads the CEO’s inbox and the shared inquiries inbox, answers routine mail and books accepted meetings on its own, and puts everything else in one evening to-do she answers a line at a time. The model drafts; code decides every send.
Ask me about what the assistant is allowed to do without asking, and how code enforces it
A weekly Claude task reads the website and podcast analytics and the Kajabi platform’s figures, and emails leadership the biggest changes, flags and suggestions. Its GA4 reports and its email go only through a no-AI n8n workflow.
Ask me about why the allow-lists live in n8n code instead of the prompt
Each month a Claude Cowork task scores new podcast guest applications, holds anyone the CEO already knows, declines weak fits on its own and sends her a digest. Leadership no longer handles applicant email.
Ask me about how it defends against instructions hidden in an application
A 343-episode archive indexed so that each week code picks an eligible episode, Claude drafts the newsletter section around a quote already on record, and n8n sends it to the editors.
Ask me about why the model never chooses the episode
Colleagues ask ten years of the podcast a question in their own Claude. Answers are built only from the transcripts, code checks every quote against its source, and a question the archive can’t answer is declined.
Ask me about why combining keyword and meaning search did worse than keyword alone
A finance operations platform that checks its own work, and the evaluation harness behind every error rate on this site. The four free tools package the same measuring for anyone to use.
Spreads financial statements, reconciles bank against ledger, matches invoices to orders and receipts, gates vendor data and scans spreadsheets for formula risks. Every engine checks its own work and sends a person only what it is unsure of; a small local model maps labels and never sees a number.
Ask me about why it won’t auto-accept a rule that has been right every time so far
My own command-line tool, built with Claude Code and still evolving: it grades any workflow export against an answer key written in advance. The headline is the silent error rate, every rate carries a range, and separate commands measure run-to-run variance and what-if thresholds. I use it regularly on new builds.
Ask me about how you test a tool whose whole job is to be right about numbers
Spec builds for real problems, measured with the same harness, bad results included. The answer keys are stress tests I wrote: 500 records each, many built to trip the agent.
Tags incoming news against a controlled vocabulary and decides which articles can publish with no editor. Rules resolve the companies, a model proposes tags with evidence, and a guard rejects anything the evidence does not support.
Ask me about why 24 test articles said zero errors and 500 said 27%
Screens law-firm advertising against attorney advertising rules. Regex catches the banned words, a model catches the implied violations, and its confidence decides what a person reviews. Since ported to tested Node.js code that matches the workflow on all 500 documents.
Ask me about why the fix went into the prompt, not the threshold
A take-home for an AI documentation product: fifteen candidate drivers of summary quality tested across 200 conversations with correction for multiple comparisons. None held up, and the write-up says so.
A scheduled Claude agent that searches three job sources every morning and keeps only genuinely AI-scoped roles; an n8n ledger then follows every application to an outcome. ~4,000 postings screened.
Decides which of a competitor’s changes is material enough to report, on a calibrated scale, in a digest built for a 90-second read.
Regression-tests a documentation prompt on real transcripts, where a softened risk or an invented diagnosis is a real harm. Same assignment as the analysis.
A one-file fitness dashboard I use daily: a calendar shaded by calorie deficit, lifts with automatic PRs and a model-versus-scale weight chart. Open source, 22 tests, no server.
Ranks every listing on a weighted model of what the buyer cares about, and flags listings priced below a set price per square foot.
I'll screen-share any build: the workflow, the test set and the results, including what didn't work.
Open to AI automation & AI engineering roles · Columbus, OH · relocation or remote