PawCheckAI reads. Rules score.
That one design decision is why PawCheck's scores are consistent and explainable: the AI never writes the number. Here's the whole pipeline, honestly explained.
Snap the front, or scan the barcode
A photo of the front of the pack is resized on your phone and sent securely to our scanning service. Barcodes work too. Photos are used to identify the product, not to build an advertising profile of you.
AI finds the ingredients
Google's Gemini AI identifies the brand and recipe. The ingredients come from our catalogue or an online lookup for that product, or from a photo of the ingredient panel. If we can't read them, we ask for a photo. Each ingredient is sorted into a fixed group (beneficial, neutral, questionable, harmful), using our own rules where we have one and the AI's classification otherwise.
The rules engine scores it
Fixed, versioned rules compute the 0–100 for your pet's profile. Named protein first? Points up. Artificial preservatives? Points down. An allergen you selected for your pet in the ingredient list? The score is 0 for that pet. No ingredient list read yet? The score is capped. Every factor is shown on screen.
Checked against FDA and CPSC recalls
Each scan is checked against the FDA and CPSC recall notices in our database, which doesn't include every recall. If a notice matches, it's shown on the result. The official notice is always the authority.
Same product. Different dogs. Different scores.
A chicken-first kibble might score 86 for a Lab with no allergies, and 0 for a dog with chicken selected as an allergy: a selected allergen in the ingredient list means Do not feed. Both owners can see which rules moved the number. Species, allergies, your strictness setting and three conditions (kidney, weight and sensitive stomach, when the nutrition panel can be read) shape the score. The app shows what it could check.
- Every factor that moved the score is shown on the result
- Same reading, same pet, same rules = same score. Rules are versioned.
- Wrong product or reading? Tap "Not right?" on the result
Tight guardrails, on purpose
Large AI models are brilliant readers and unreliable narrators. So PawCheck uses the model to identify and read, never to score:
- Allowed: identifying the product, reading or finding its ingredients, sorting ingredients into fixed groups
- Not allowed: writing the score
- Recall checks: against the FDA and CPSC recall notices in our database, which doesn't include every recall
- When unsure: if we can't read the ingredients, we ask for a photo
