Trust is an engineering problem. Here is how we solved it.

The hard NOs

Leo never diagnoses a condition, never prescribes or adjusts a dose, and never declares something "safe" or "dangerous". He helps you understand what's going on and points you to your vet when it matters. This is enforced, not requested: every reply passes output validation that blocks diagnosis-shaped or verdict-shaped text before you see it, and every incoming message is screened for prompt-injection and jailbreak attempts before Leo reads it.

Emergencies bypass the AI

When a message looks like an emergency ("she ate chocolate", "he collapsed", seizures, bleeding, poisons, shock, trauma), no AI runs at all. A rule-based detector fires first on every single message and returns instant first-aid steps read from reviewed data, never model-generated, plus who to call: your own vet, preferring one you've flagged for emergencies, and the correct poison-control line for your region.

• Missing a real emergency is the failure we optimize against, so the detector is tuned to err towards flagging.
• Hypotheticals like "can dogs eat chocolate?" don't false-alarm.
• Contact numbers come from reviewed data, never from the model. Leo does not invent phone numbers.

Veterinarians review the clinical content

First-aid instructions, canned answers, emergency contacts and the seasonal risk calendars all go through a veterinarian review workflow. Approvals are bound to a content hash, so the approved text is exactly the text served, and review decisions are append-only, which makes sign-offs tamper-evident. If the content changes, the approval no longer applies.

Review is still in progress, and we would rather tell you that than imply otherwise. The regional risk calendars for the United States, United Kingdom and India are vet-approved today. Other clinical content, including the wellness-plan templates and the toxin guidance, is still awaiting sign-off, and the app labels it as draft guidance on the screen where you read it. Serving reviewed and unreviewed content is a deliberate decision; serving them indistinguishably is not.

Deterministic where it matters

Questions about your own records ("what vaccines has Mango had?") are answered straight from the database with zero AI generation. The risk profile is computed from versioned catalog data: vet-approved country climate files, a pathogen registry, and per-product protection durations. No AI generates clinical values; Leo only narrates them. The app and Leo run the same evaluator logic, verified against shared golden test cases, so the card on screen and the answer in chat do not disagree.

Honesty about precision

If Maggie only knows your country, she gives country-safe advice rather than pretending to zone-level precision. If a dose history is incomplete, she gives no coverage verdict at all. Saying less when we know less is a design rule.

Wellness plans have gates, not vibes

Weight and fitness plans are assembled only from a fixed set of reviewed templates, never generated per pet, and matched to your dog's real weight, WSAVA body-condition score, breed, age and size class. Safety gates run first: severe overweight, underweight, rapid weight loss and puppies are routed to "see your vet" instead of being served a plan. A plan auto-pauses if weight drops too fast. When the data says see a vet, Maggie refuses to guess. These templates are the content still awaiting vet sign-off described above, and they carry the draft-guidance label in the app.

Your household, and only your household

Every request is authorized server-side against your household. Leo can only ever see pets that belong to you and the family members you've invited with your 6-letter code. There is no path to anyone else's data, and no request skips the check.

Privacy and erasure

Chat transcripts are not persisted long-term. Personal information is detected and handled by type; actual values are never written to logs. Account deletion has a 30-day grace period, after which server-side erasure sweeps remove your data. The flow is built to support the erasure rights set out in GDPR and India's DPDP Act. Free and Pro users get the same top-tier AI model; safety is never a paid feature, and emergency help is free for everyone, always.