// AI

A point of view, not a feature list.

In regulated, expert-heavy products, AI's job isn't to replace the expert. It's to remove the friction between the expert's judgment and the action. I build small tools to keep that opinion honest.

// THE TAKE

AI earns autonomy the same way a new hire does: narrow scope, a track record, then wider latitude. Anything else is a demo, not a system.

The teams that get this right embed AI inside the existing workflow (the system of record, the handoff, the queue) instead of bolting on a separate app the expert has to context-switch into. And they let trust compound in increments: tight scope first, human override always live, autonomy widening only as the system proves itself in that lane. Skip the increments and the failure mode shows up in production instead of in a demo, and in healthcare, the failure mode is the product.

I hold that view because I've built the workflow platforms where embedding, not add-on features, was the unlock. And Keystone, below, is that same trust-compounding logic applied to how I build with AI myself: narrow gate, human call, then the next stage opens.

// BUILDS

Small builds I ship myself to pressure-test what AI can actually do, and to keep my engineering, product, strategy, and venture-investing instincts sharp while I learn it firsthand.

APPLIED AI · VISION + GENERATION · LIVE ON iOS

Picture-to-Plate

An image-to-recipe app, live on the App Store as an MVP. Photograph what's in your fridge or pantry, get a recipe back. I built it to see where vision models are reliable and where they hallucinate, and what that means for product decisions in domains where the failure mode matters. It's also where I'm capturing real usage data on where those failure modes actually show up.

What it taught me: the model already handles the vision and the generation. What it can't do alone is remember your pantry and weigh that against what you're in the mood for, that's the workflow layer, and the payoff is a curated shortlist (3 almost-there, 3 a short trip, 3 worth planning around), not endless recipes.

Get it on the App Store →

APPLIED AI · PRODUCT ORCHESTRATION LAYER

Keystone AI

A human-in-the-loop orchestration layer for taking an idea to a validated MVP. Skills and agents do the ground work at each of six stages; the judgment, the constraints, and the go or no-go call stay with the expert in the loop, here, a product manager. Nothing advances without that call.

What it taught me: agents are confidently wrong often enough that the gates aren't overhead, they're the actual product-management contribution. Remove one to move faster and you ship something broken.

View on GitHub →

NEXT UP · TRENDSHELF

Turning restock guesswork into an evidenced call for independent online sellers.