Fractional product lead / design-led AI builder, 25 years shipping commercial work. I help non-technical founders and operators decide what to build, then ship it as production software. Not prototypes, not demos.
Recent ships: Steadcast (steadcast.co), a native macOS podcast app in the Mac App Store, plus its production AI backend: queue-based Whisper transcription on a self-hosted GPU, eval gates that block bad model output before it persists, mixture-of-models routing, cost ceilings, monitoring. Currently building an on-device Apple Intelligence macOS app that fills distributor and retailer paperwork for CPG brands.
Background: Apple Design Award winner (2003, Tubatomic; commercial work for Warner Bros, HGTV, Turner). Ran Hoff & Pepper hot sauce for a decade: 1M bottles, ~5,000 retail doors, Hot Ones Season 16. I've sat on both sides, operator with a P&L and the person shipping the software.
Best fit: AI features that need to actually work, founders burned by agencies, CPG/DTC operators who want a builder who has run a real business.
25-year design-led product builder. Apple Design Award winner (2003, Tubatomic, commercial work for Warner Bros, HGTV, Turner). Ran Hoff & Pepper hot sauce 10 years (1M bottles, 5,000 retail doors, Hot Ones S16) before winding it down. Past 12 months: shipped Steadcast (native macOS podcast app, live in the Mac App Store at https://steadcast.co) and its Knowledge Engine: queue-based Whisper on self-hosted RTX 3060 (17.9× realtime), 410-rule editorial voice guide, Promptfoo eval gates blocking bad output before it persists, mixture-of-models routing (Llama 3.3 for structured extraction, GLM-5.1 for prose), cost ceilings, monitoring, soak windows.
Best fit: AI features that need to actually work in production. Non-technical founders burned by agencies. CPG/DTC operators who want an operator-builder, not a consultant. Fractional Product Lead engagements.
The thing that made Roblox actually work for kids wasn't the editor, it was that a kid could send a link and their friend would be playing 10 seconds later. S&box, Hytale, Luanti are all fun but they're all installs, which kills the sharing loop the moment you have one friend on an unsupported platform or a school laptop.
The closest thing to the Roblox distribution model is browser games. I work on browser-based game stuff and the hard part is never authoring, it's the last mile: corporate networks, WebSocket proxies, "my friend has a Chromebook," etc. Godot has a web export now that's genuinely usable for small multiplayer stuff, and it's free. Not as polished as Roblox Studio but the zero-install property is the whole ballgame for kids sharing with friends.
Yeah, this is the part of building browser games that's quietly gotten harder in the last year. I've been working on a Quake III conversion that runs in the browser and I keep bumping into versions of the same problem. Anything the client knows, an agent knows in about five minutes now.
The weird thing is the fix (make the server authoritative) is 90s MMO wisdom, it just used to not be worth the effort for a small game with no money on the line. Now the cost of a bot attempt is basically zero so every game inherits the threat model of a game that does have money on the line, whether you want it or not.
The leaderboard-split thing OP ended up doing is probably the right call for a lot of these. Fighting it is a losing battle when one person with Claude and a weekend can out-iterate you.
Remote: Yes (US)
Willing to relocate: No
Technologies: Swift, SwiftUI, macOS, Apple Intelligence (on-device), Python, Postgres, LLM pipelines (Llama, Whisper, GLM), Promptfoo evals, product design, front-end
Resume/CV: doubleabattery.com/work-with-me (full resume on request)
Email: [email protected]
Fractional product lead / design-led AI builder, 25 years shipping commercial work. I help non-technical founders and operators decide what to build, then ship it as production software. Not prototypes, not demos.
Recent ships: Steadcast (steadcast.co), a native macOS podcast app in the Mac App Store, plus its production AI backend: queue-based Whisper transcription on a self-hosted GPU, eval gates that block bad model output before it persists, mixture-of-models routing, cost ceilings, monitoring. Currently building an on-device Apple Intelligence macOS app that fills distributor and retailer paperwork for CPG brands.
Background: Apple Design Award winner (2003, Tubatomic; commercial work for Warner Bros, HGTV, Turner). Ran Hoff & Pepper hot sauce for a decade: 1M bottles, ~5,000 retail doors, Hot Ones Season 16. I've sat on both sides, operator with a P&L and the person shipping the software.
Best fit: AI features that need to actually work, founders burned by agencies, CPG/DTC operators who want a builder who has run a real business.