Each skill, and where to see it.

Designing agents that run in loops, hand work to each other, and keep their state in files a person can review.

• A content operation run by AI agent loops
• Evals for AI-generated output, and testing the tests
• Context engineering: a doctrine every writing agent loads

Writing the standards, maps, and workflows a model loads, so its output is reliable at volume rather than good once.

• Context engineering: a doctrine every writing agent loads
• A content operation run by AI agent loops

Checks that grade real output on every change, an audit that proves the checks can fail, and endpoints that report what is actually wired.

• Evals for AI-generated output, and testing the tests
• A health endpoint that told the truth about the funnel

Agents propose; a person approves. Every change is a reviewable pull request, and known failure classes are blocked by code.

• A content operation run by AI agent loops
• Evals for AI-generated output, and testing the tests

An internal agent with real users, a measured before and after, and an owner who is not its builder.

Now building: A recurring reporting narrative agent for a marketing team, with a weekly usage and accuracy log.

This site: Next.js 16 and React 19, roughly 840 TypeScript files, and the build tooling around it.

• A content operation run by AI agent loops

Typed, tested Python for agent services and data work.

Now building: A lead research and routing agent written in Python with pydantic models and a pytest suite.

One modelled source of truth for marketing data, built in a warehouse and versioned in dbt.

Now building: A marketing data model joining paid media, web analytics, and point-of-sale data by location.

Measuring each layer of a funnel, ruling causes out explicitly, and stating what the data does not show.

• Why a 995-page site drew zero search traffic
• A health endpoint that told the truth about the funnel

Forms, webhooks, email, booking, analytics, and payments wired together so data arrives where it is needed and failures are visible.

• A health endpoint that told the truth about the funnel

Inbound leads researched, scored, and routed to the right person automatically, with a review queue for low-confidence calls.

Now building: The same Python agent as above, writing to a CRM with an eval set of labelled leads.

Tests with a stated hypothesis, a sample size calculated in advance, and a readout someone can act on.

Now building: A local-marketing test across matched store groups, analysed as a difference-in-differences.

Write-ups that lead with the decision, carry the evidence, and say plainly what is not yet known.

• Why a 995-page site drew zero search traffic

Designing for use from the first day, and measuring whether people still use the tool a month later.

Now building: Weekly active use and time saved, tracked for every internal tool shipped this quarter.

Deciding what not to build, testing ideas on paper before spending on them, and retiring what does not earn its place.

• An offering graph to choose one thing to sell
• Why a 995-page site drew zero search traffic

Work delivered as small, stacked pull requests a reviewer can accept one at a time, with retractions flagged across the stack.

• A content operation run by AI agent loops

Choosing what to sell, to whom, and on what pricing model, from measured demand rather than preference.

• An offering graph to choose one thing to sell
• Why a 995-page site drew zero search traffic

Day-to-day operating experience inside a multi-location specialty retailer: store-level reporting, local marketing, and the gap between dashboard and store floor.

• An offering graph to choose one thing to sell

A client outcome counted in dollars, verified by the client, and published with permission.

Now building: Missed-call recovery for home-service businesses, priced on the bookings it produces.