Projects

I lead product, and I still build. Each project here started from a real problem, and each one shows how I decide what to build. Most of the code is on GitHub.

Product work

AI trust research

AI summaries that can be steered

The problem: AI summaries can be steered by text that a human reader never sees.

What I did: built tools to test it, and wrote up both sides: how a brand can make sure AI summaries describe it accurately (Summary Ranking Optimization), and how hidden signals can mislead them (Summary Response Manipulation).

What came of it: published research, and a flaw in Google’s Gemini that I reported and Google accepted as a P2/S2 bug.

Summarizer on GitHub · The Gemini disclosure · All research

AI product design

Radically Honest

The problem: people can’t see what an AI assistant is for, how it behaves, or where its limits are.

What I did: designed a GPT that states its purpose, behavior and limitations up front, and made the whole system open to audit.

Radically Honest on GitHub

Developer tool

Reflex

A skill system for Claude that scales to 78 modules, so one assistant can take on many kinds of work without its instructions breaking down.

How Reflex works

Agentic AI

Agentic workflow starter

A starting point for multi-agent workflows with CrewAI. A safe kickoff lets a run finish cleanly when one step fails, and a custom tool gathers brand images for video mood boards.

crewai-agentic-starter on GitHub

Also on GitHub

HRmageddon

A modern web remake of the 2009 Adult Swim Flash game where rival departments wage cubicle warfare. It’s a turn-based strategy game with a data-driven map system and a full single-player mode against AI.

HRmageddon on GitHub

Hiring Manifesto

As an industry we’ve written manifestos for how we build, ship and scale, but not for how we hire. This is a manifesto and practical framework for humane, values-driven hiring, because hiring is the first product a company ships.

HiringManifesto on GitHub