PawNexus
How PawNexus works
Every recommendation is grounded in verified data — not in what an AI happens to remember. Here's the pipeline that gets you trustworthy answers for dog and cat care decisions.
The problem with generic AI for pet care
Generic AI chatbots will confidently tell you the wrong thing about pet insurance plans, food ingredients, breed health risks, or claim eligibility. They hallucinate carrier names, invent policy clauses, and mix up species-specific facts. For decisions involving real money or your pet's safety, "confidently wrong" is dangerous.
PawNexus is designed differently. Every answer is grounded — meaning the AI is only allowed to recommend things that exist in our curated source-of-truth databases, and every recommendation is traceable back to a verifiable record.
The grounding layer
PawNexus combines large language models (Google Gemini) with three deterministic data sources:
- PetBreedManifest — a curated database of dog and cat breed traits, health risks, energy levels, grooming needs, and size categories. When PawNexus reasons about a breed, it pulls from this record, not from LLM memory.
- Parsed insurance policy documents — actual carrier policy PDFs are processed and indexed with page-level citations. Every plan recommendation comes with evidence pointing to the specific policy clause.
- Licensed provider directories — vet, trainer, and boarding listings come from Google Places enriched with certification registries (AAHA, AAFP, CCPDT, IAABC, KPA-CTP).
How a recommendation is built
- Intake. You describe your situation in natural language — your pet, your home, your priorities. PawNexus extracts structured fields (species, breed, age, ZIP, medical conditions, preferences) into a state object called PetScene.
- Layered fallback extraction. If the first extraction pass misses a field, a second layer queries the breed manifest, and a third layer asks the LLM a targeted, permissive follow-up. This handles edge cases like "Chihuahua doesn't imply dog" or international ZIPs leaking through.
- Grounded retrieval. PawNexus pulls candidate records from the relevant data source — insurance plans available in your state, vets near your ZIP, trainers with matching certifications — without inventing anything.
- Deterministic scoring. Candidates are ranked using transparent, rule-based scoring on fit signals (coverage priorities, breed health risks, certifications, distance).
- Evidence-required output. The LLM is constrained to only describe candidates that survived retrieval and scoring. If a claim cannot be sourced, it is dropped — no hallucinations allowed.
Why this matters for you
The grounding pipeline is the difference between "an AI guessed" and "this is what the actual policy says." Practical examples:
- If PawNexus tells you a plan covers hip dysplasia, that recommendation traces back to a specific page in the carrier's policy document.
- If PawNexus suggests a trainer, the certification badge shown is sourced from the certifying body's registry — not an LLM guess.
- If PawNexus warns about an ingredient in your pet food, that flag comes from breed-specific safety records, not generic advice.
Patent posture
Certain aspects of the PawNexus architecture — including the grounded AI recommendation pipeline, accumulative natural-language state compilation, and deterministic evaluation gating — are the subject of pending patent applications.