Four AI products, four different AI architectures

Why

AI changes what one person can ship — and it changes what "good AI product judgment" means. Reading about that is one thing. To pressure-test it, I set out to ship four complete consumer products, each built on a deliberately different AI architecture: cloud LLM, a cloud model with an on-device fallback, fully on-device, and no model at all. Every one has its own brand, website, and users' data to protect.

Cleared — an opinionated AI to-do app

clearedfirst.com · "Clear what matters. First." · live on the App Store

Most to-do apps are lists; Cleared has an opinion. Say a goal and AI plans the steps. Your tasks then compete — you judge them head-to-head in a bracket, or AI referees — producing a ranked tier list with one task spotlighted as clear next. Work in order until the app says "All clear."

• The AI bet — cloud LLM economics as a product constraint. The entire ranking runs in a single cost-optimized Gemini request, and app attestation ensures only the real app can spend model quota. Latency and unit cost were designed, not discovered.
• The agent bet — the app is also a tool surface. Cleared ships a remote MCP server, so Claude can read the list in priority order, file tasks into tiers, and run the same prioritization the app runs. It signs in with OAuth 2.1 against the app's own accounts, and writes go through the same rules the UI does — so anything Claude does is undoable in Cleared. Connecting is one URL and one approval: no API keys, nothing to install.
• React 19, TypeScript, Tailwind, Firebase (auth, sync, App Check), Gemini, and a remote MCP server on Cloud Functions — shipped as a PWA and, via Capacitor, an iOS app on the App Store.

IronProof — proof over motivation

ironproof.org · "The scale lies. Proof doesn't." · live on the App Store

A fitness progress tracker built on a simple insight: photos reveal what scale weight can't. IronProof turns progress photos into a timeline of transformation — weight and measurements overlaid on every shot, dates read automatically from photo metadata, synced across devices.

• The bet — breadth of shipping. One codebase runs on the web, iOS, Android, and desktop. It went through Apple's review and is live on the App Store.
• The hard parts were unglamorous product engineering: HEIC photo decoding, EXIF date extraction, offline-first storage with cross-device sync.

Grava — calorie counting, without the counting

gravacal.com · "Snap. Confirm. Eat." · live on the App Store

A native iPhone calorie tracker that removes the logging that kills every diet app: photograph a meal, a menu, a nutrition label, or a delivery screenshot, and AI estimates calories and macros. Estimates come back as honest ranges, and the app biases toward the top, middle, or bottom of the range depending on whether you're cutting, maintaining, or bulking — protecting the user's goal instead of faking precision.

• The AI bet — two models behind one estimator. Built in SwiftUI. Estimates run on Gemini by default, because accuracy is what makes a calorie estimate worth trusting; when the request fails, the quota is spent, or the phone is offline, the app falls back to Apple Intelligence on device — the Vision framework for food recognition and OCR, plus the on-device foundation model, calibrated with serving-size anchors the smaller model needs. The UI names which model answered, so a degraded estimate is never passed off as the best one.

Delecto — culling a shoot, tournament-style

delecto.app · "600 frames. Four keepers. One bracket." · live on the Mac App Store

A native Mac app for photographers. Star ratings ask a question nobody can answer at a glance — three stars or four? Delecto asks the one you can: which of these two is better? It shuffles a shoot into a single-elimination bracket and shows two photos at a time. Winners advance round by round, and you can stop whenever you're down to the handful you wanted. Every pick is saved the moment you make it, and no file moves until you export.

• The AI bet — fully on-device, designed around what the model can't do. Optional Apple Intelligence culling decides every match on the Mac itself: no networking code, no account, no API key, no cost per decision. On macOS 26 the on-device model reads text, not pixels, so each photo is first measured with Apple's Vision framework — aesthetics, sharpness, lens smudge, clipped exposure, faces, horizon tilt — and the model judges the two sets of measurements against criteria the user can rewrite. Every verdict carries a one-line reason; when the model's guardrails decline a match, it's decided on quality scores alone and labeled that way.
• Honest about speed. A verdict takes about 8 to 20 seconds on an M1 Pro, so AI culling is framed as an unattended run — start it, make coffee, come back to a decided bracket — and Suggest shows which photo the AI would pick, and why, without deciding for you.
• SwiftUI, Vision, Foundation Models, and ImageIO, which decodes HEIC and camera RAW natively — sandboxed, with zero third-party dependencies. Semantic-release cuts versions; Xcode Cloud builds and uploads to the Mac App Store.

What building all four taught me

• Architecture is a product decision. Cloud models buy capability and cost you unit economics; on-device models buy privacy and offline reach and cost you accuracy. Grava started fully on-device and moved to cloud-first once testing showed the small model's calorie estimates weren't good enough to trust — keeping it as the fallback rather than shipping "works offline" as a promise the accuracy couldn't cover. Delecto made the opposite call on the same tradeoff: its promise is that photos never leave the Mac, and the hard quality calls are measurements Vision makes reliably, so the small model only has to add taste and an explanation. Choosing per product — and revising the choice when the evidence changes — is the actual skill.
• AI compresses build time, not judgment time. The models wrote a lot of code. Deciding what each product refuses to do — Cleared's single-request budget, Grava's ranges instead of false precision, Delecto never deciding a match without saying why — was still the work.
• Shipping is the credential. Brands, websites, App Store review, auth, sync, data protection — the last 20% is where product managers earn the title.