Yummster
Live · closed beta"What's for dinner?" is a question millions of households in Mexico answer every single day, usually while staring into a fridge. Yummster answers it from what's actually in that fridge.
The problem nobody fixed properly
Recipe apps assume you'll go shopping. Pantry apps assume you'll do inventory homework. Real people do neither. The pantry-tracker category is a graveyard of apps that asked users to become clerks, and recipe apps keep recommending dishes you can't make with what you have.
The insight Yummster is built on: the pantry should stay accurate as a byproduct of cooking, not as a chore. You cook, you confirm what got used in one tap, and the app quietly learns your kitchen. No barcode scanning. No expiry spreadsheets.
Decisions that shaped the product
the audits caught 3 fail-open allergen bugs before any tester cooked. worth every token- Three states, never quantities. Every ingredient is In Stock, Running Low, or Out. Quantities are what kill pantry apps, so they were banned from day one. Status changes are stored as an append-only event log, which doubles as training data for the recommendation engine.
- Safety is not a ranking problem. Allergens and dietary restrictions run as deterministic hard filters before any embedding-based ranking ever sees a recipe. The gate fails closed: if the app can't verify your restrictions, it shows nothing rather than guessing. Adversarial audits caught three paths where the filter could fail open. All three were fixed and covered with end-to-end tests before any beta tester touched the app.
- Every suggestion explains itself. Recommendations carry a true "why this" generated from the same signals that ranked them. If the app can't honestly explain a suggestion, it doesn't fabricate a reason.
- Local first, really. The catalog is authored in natural Mexican Spanish, not translated American recipes. When a batch of localized content drifted toward translation patterns, it was rewritten rather than shipped.
Building a product org of one
Yummster is designed and shipped by one person, me, running AI coding agents the way a product leader runs a team: written specs, locked decisions, adversarial review before merge, and a canonical project canvas that outranks anyone's memory, including mine. The design system has an automated gate that fails the build if code tokens drift from the design source. Cross-model audits review the work my primary tools produce, and they've caught real bugs, including the allergen fail-opens above.
The result is a real shipping cadence without a payroll: a live app on device with the full cook loop working, telemetry and crash reporting wired, store builds passing on both platforms, and a bilingual marketing site with a live waitlist, all while the product spec, schema, and legal surfaces stayed documented enough that a new engineer could onboard from the files alone.
The business underneath the app
The consumer product is a subscription cooking app. The venture story is bigger: every cook session teaches Yummster what a household has, likes, and is about to run out of. That first-party pantry and taste graph is a wedge into LATAM grocery commerce, converting missing-ingredient intent into baskets and retailer partnerships. Two lines were drawn early and permanently: allergy data is never used for targeting or insights, and no user-level data is ever sold or brokered. That's both an ethical position and a durable trust advantage in a market that has been burned before.
Pricing went through a researched ladder with trial mechanics and unit-economics modeling (lifetime value against cost to serve, with a hard infrastructure ceiling), so the freemium boundary is a business decision, not a guess.
Where it stands
- Full MVP cook loop live in closed beta: onboarding, discovery, cook mode, post-cook pantry confirmation, shopping lists.
- Marketing site live at yummster.app with a working waitlist, privacy notice, and bilingual SEO surface.
- First on-device smoke test passed end to end, analytics flowing, crash reporting armed and quiet.
- Next milestones: beta cohort onboarding and store launch rails.
What this project says about me
I didn't build a recipe app with AI sprinkled on top. I found the structural reason a whole category keeps failing, designed around it, and shipped the thing solo to a live beta with the safety engineering of a much larger team. If you want to know how I'd treat your product, this is the honest sample.