PawCheckAI reads. Rules score.
That one design decision is why PawCheck's scores are consistent, explainable and impossible to hallucinate. Here's the whole pipeline, honestly explained.
Your photo, compressed & sent
The label photo is resized on your phone and sent securely to our cloud scanner. Photos are used to identify the product — not to build an advertising profile of you.
Vision AI transcribes — verbatim
A vision model reads the packaging under strict rules: transcribe only what's visible, in printed order. Illegible field? It returns "unknown" — never a guess. It also classifies each ingredient against a fixed taxonomy (beneficial, neutral, questionable, harmful).
The rules engine scores it
A deterministic, versioned rules engine computes the 0–100 against your dog's profile. Named protein first? Points up. Artificial preservatives? Points down. Your dog's allergen in the list? Big penalty — with the exact number shown on screen.
Recalls joined from the database
Recall status comes exclusively from a recall database fed by the FDA's animal & veterinary recall feed, each with a source link. The AI has no ability to add a recall to your screen.
Same product. Different dogs. Different scores.
A chicken-first kibble might score 86 for a healthy adult Lab and 58 for a senior Frenchie with a chicken allergy — and both dogs' owners can see exactly which rules moved the number. That's the quiz working for you: breed, age, weight, allergies, conditions and goals all become scoring inputs.
- Every factor is visible: "+5 named protein first", "−18 contains chicken (your dog's allergen)"
- Scores only change when the rules version changes — and the app tells you when that happens
- Re-scan the same bag next week: identical score, guaranteed
Tight guardrails, on purpose
Large AI models are brilliant readers and unreliable narrators. So PawCheck only ever uses the model as a reader:
- Allowed: transcribing the label, classifying ingredients against a fixed list
- Not allowed: inventing ingredients, authoring scores, claiming recalls
- When unsure: low confidence means "retake the photo," never a best guess
See it work on a real shelf.
Or try the interactive demo first.