I build AI systems.

I built a client-reporting portal for a division of the firm.

It runs on an append-only fact store: nothing is ever overwritten, every number keeps its provenance (a record of where it came from), and each client who signs in sees only their own report. Each account statement is logged once, any figure the AI pulls out without a matching passage in the source document is thrown away, and each fact is saved with its source, date and method.

A newer fact replaces an older one without erasing it. Automated code rules, known as lint rules, refuse any number typed in by hand, and a report that does not add up cannot publish. A number that could be read two ways, such as 1,234 or 1.234, goes to a person. When two sources disagree, it never picks one on its own; it records the conflict and sends it to a person. A client can click any number to see the passage, source and dates behind it.

It has been built and tested only on made-up demo data: no real client report has been published, and the sign-off step has never been used. The code went through a security review, and every issue it found was fixed.

An evaluator-optimizer pipeline (one AI drafts, others grade, and it rewrites) that turns an interview into social media posts.

The interview keeps asking until it has what each post needs, such as a hard number or a lesson from a mistake, and the author’s exact answers become the source file, with every claim tagged to the answer it came from. The system outlines first, writes several drafts and keeps the best one.

The deterministic check blocks any number, name or quote that is not in the source file, and a second check stops any claim on the firm’s off-limits list. The judges are an AI editor, writing judges, a detector for AI-sounding text, a Reader judge that asks whether a reader would reply, and a compliance judge who can veto a draft. The rewrite loop fixes the biggest problems first, keeps what the judges liked and stops to ask the author once drafts stop improving.

It is in use now. When I tested the judges against engagement on a set of posts, their average writing score did not predict it, and only the Reader judge’s yes or no did. Each post now has to name its goal, and some must end on a question readers answer about themselves. In one test, a judge’s veto on made-up facts caught an interview exchange that never happened.

It runs inside Claude Code, an AI coding tool, as a step-by-step build process: each stage must pass its checks before the next begins, and every check cites the research behind it.

It starts with an interview that produces a brief and a list of every claim the site will make, and an unconfirmed claim holds its section back. It studies the industry and the best sites in the field, the owner picks a design direction, and it plans each section’s purpose, heading and facts before building one section at a time.

The checks include every claim traced to its source, no placeholder text, nothing spilling off the screen at any width tested, measured Core Web Vitals (Google’s page-speed measures), automated accessibility tests and a working contact form. It publishes a preview, then the live site, and runs the checks again on the live address, though no recorded run has yet passed that last step.

I built it because every site I made meant working out the same process again, and a step-by-step process makes the minimum quality the same every time. The first version lost the blind design comparisons because it left placeholder text, buried the contact form and repeated one section layout. I added a check for each, and the next version won every judged comparison. Later, a fact-check pass deleted a product’s own name from a test page, so the brief’s lines stating who the business is are now facts no checker can remove. This site was built with it.

Two marketing sites for EG Capital: a redesign of the firm’s main site and a site for one of its divisions, each built from a list of sourced facts.

Before writing the main site’s copy, I studied its current site, the competition and a set of reference sites, page by page. I built the division’s site in plain HTML, CSS and JavaScript and rebuilt the main site on Astro, a toolkit for building fast websites, with a page for each audience.

On the division’s site, teams of AI reviewers checked brand, wording against investment marketing rules, accessibility, speed and design, with an adversarial verifier trying to prove them wrong. The main site passes a strict code validator and scores near the maximum on Lighthouse, Google’s site audit. The main site carries the division’s legal disclosure word for word.

The firm’s site went live in November 2026, and the division’s site was taken offline in October 2026.

The weekly research system that ran inside the podcast system above: it found people who fit a written ideal guest profile and drafted invitations.

It found candidates in public SEC filings and on rival shows, then scored each on fit, recent news and how reachable they were, with older news counting for less. A top rating needed two of three sources: a deal announcement, an SEC filing or a state business record.

The compliance check ran before any writing, and only then did it draft a short series of emails. A last check held any draft with a claim it could not confirm, and invitations waited as drafts (later in Gmail).

The research-first rule also cleared the pile of flagged people. One drafting pass sorted people by how strong their news was and ignored the recorded "not a fit" marks, so it drafted invitations to people already ruled out; every pass now checks that mark first.

An orchestrator for multi-agent work: it puts a fresh team of AI specialists on one problem, briefs each from published methods and brings their combined work back to me to decide.

Each run starts from a brief: the problem in my words, what is already settled, and measurable signs of success. It researches how the strongest organizations staff a problem like this, then gives each specialist one question and a brief with cited sources.

The specialists work at the same time, the lead records a decision on each disagreement, and the outside reviewer checks the result. It ends on a report that leads with the outcome and puts my open decisions to me as yes-or-no questions.

Every specialist is a sourced brief plus written instructions, never a persona, because my own standard for working with AI agents does not allow role-play. When the standard test scored every installed skill zero, I wrote my own test of whether each one activates when called. I have used it since October on problems that needed several specialists. Briefs on some topics came back thin because most of the best sources were behind paywalls or blocked.