Five Across
A multiplayer bingo game built in eight days for a nine-night cruise, then operated live at sea—where real players exposed what the architecture could and couldn't prove.
Products—and the decisions behind them
Every one of these began as a real problem—mine, or one I watched up close. The decisions are the part worth reading: what got built, what got refused, and what the evidence said once it was running. Each card says where the project stands, and why. I build with AI agents under a review system I designed and documented here. The agents are leverage; the product decisions are the point.
Mergepath
Agents could recite the no-direct-push rule and skipped it anyway, so enforcement became gates that refuse—then an audit found eight of ten repos enforcing zero of five checks.
Override
Broadway financing turned from static documents into an interactive model of the deal—built as an AI coding exercise, and still untested on a real capitalization.
Device Source of Truth
Five fragmented data feeds became one device registry, with AI drafting the heaviest intake and humans deciding what becomes authoritative.
Matchline
Paused with better instruments and the same 51.3% extraction accuracy against an 80% bar: the last two working weeks fixed the ruler, not the product.
Swipe Watch
A swipe-first streaming recommender whose prototype accidentally faked the payoff it was meant to test—making reactions to the demo impossible to separate from reactions to the idea.
Friends & Family Billing
Shared billing for nine people that handles the explaining and asking without recipient accounts. Its three biggest failures shared one gap: the model was right before the interface was.