Under the hood

AI 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.

STAGE 1

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 โ€” not to build an advertising profile of you.

STAGE 2

Vision AI transcribes โ€” verbatim

A vision model reads the packaging under strict rules: transcribe only what's visible, in printed order. Illegible field? It returns "unknown" โ€” never a guess. It also classifies each ingredient against a fixed taxonomy (beneficial, neutral, questionable, harmful).

STAGE 3

The rules engine scores it

A deterministic, versioned rules engine computes the 0โ€“100 against your pet's profile. Named protein first? Points up. Artificial preservatives? Points down. Your pet's allergen in the list? Big penalty โ€” with the exact number shown on screen.

STAGE 4

Every scan checked for recalls

Recall status comes from a curated list of FDA animal & veterinary recalls, plus a live check using Google's Gemini AI with Google Search for recall notices about that exact product. A recall only counts if it names that brand and product line, and it's shown with its source.

Personalization that's math, not magic

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 โ€” 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", "โˆ’18 contains chicken (your pet's allergen)"
  • Scores only change when the rules version changes โ€” and the app tells you when that happens
  • Re-scan the same bag next week: identical score, guaranteed
The allergy step of the dog profile quiz
What the AI is allowed to do

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 or authoring scores
  • Recall checks: a live search that has to match this exact brand and product line and name its source
  • When unsure: low confidence means "retake the photo," never a best guess
The PawCheck scanner

See it work on a real shelf.

Or try the interactive demo first.

iPhone only for now โ€” Android coming soon