DECISION 1 The AI decision that mattered was saying no to AI
An AI-era product, built AI-first, whose core engine deliberately contains no LLM.
The obvious 2026 build was "wrap an LLM, let it recommend meals." I chose deterministic vet-standard rules (AAFCO/FEDIAF-based) instead, for three reasons that compound in a health-adjacent product:
- Explainability: every recommendation traces to a published nutritional standard, not a sampled token.
- Testability: deterministic rules can be pinned by assertions; a model's output can only be spot-checked.
- Unit economics: zero marginal cost per recommendation. No API bill scaling with users.
Where AI does belong, I use it hard: Claude Code as the build team across product, engine, and infrastructure. The judgment call is knowing which side of that line each component sits on.
One boundary is absolute: CatCulate informs feeding decisions, it is not veterinary advice, and the product is designed to never cross that line. Health-adjacent means the safety bar is higher, not that the app plays doctor.
DECISION 2 Safety culture without a safety team
A solo founder has no QA department, so I built one out of software:
- 600+ assertion evaluation suite across 45 golden scenarios: real cat profiles with known-correct nutritional answers, run before every release.
- Adversarial fuzzer that generates hostile inputs hunting for recommendations that violate safety constraints.
- Safety gates across 11 chronic conditions (kidney disease gets special handling), plus reproductive-state and breed-aware rules.
- Release gating: nothing ships unless the whole battery passes.
DECISION 3 Trust features nobody asked for, that everybody needed
- Label-claim verification: a data-integrity layer that independently checks commercial cat food "complete and balanced" claims against the product's own guaranteed-analysis data. The app doesn't take the label's word for it.
- FDA recall monitoring: an automated pipeline watching recall feeds, with human-in-the-loop review before user-facing alerts. Cat parents find out when their brand is recalled.
Operator’s rules applied here
- Debug the denominator: the eval suite exists because a solo founder has no one else to catch a wrong number before a cat's bowl does.
- Segment before you average: 11 chronic-condition gates instead of one "average cat" recommendation path.
- Walk the journey yourself: I am the app's first user; my two cats are its first customers.
READOUT Where it stands
Builder readout — product · engine · evals · store · domain · support: every layer of this app has exactly one person on call. That scope is the point of the record.
Small numbers, honest numbers. The point of this record isn't scale, it's demonstrated range: the same person who ran retention at 300M MAU can also ship the entire aircraft alone, with the safety culture of a much bigger team.
My cats remain magnificently unimpressed. The humans using it are happier.