Purpose-built AI for dogs and cats.
Not a wrapper. Not generic.

BossPaws AI Engine

A 2,055-line multi-model orchestrator.

The BossPaws AI Engine is the central nervous system of the AI layer — a 2,055-line orchestrator that routes every request to the right model based on input type. Text conversations are reasoned over by Claude Sonnet. Videos pass through Gemini Flash for exhaustive observation before Sonnet writes the clinical assessment. Photos are classified and analyzed directly by Sonnet Vision. PDFs are rendered to images by CloudConvert, then read by Sonnet Vision. Background memory and curation runs on Claude Haiku.

Why it matters: Before a single token is generated, the engine performs parallel fetches — BossRadar health context, curated memory profile, and owner name — so every response is contextually aware of the pet’s breed, age, weight, conditions, medications, allergies and personality. Emergency detection scans every response for critical keywords (bleeding, unconscious, difficulty breathing) and triggers a 3-tier severity alert. Response length adapts dynamically: longer for health queries, tighter for casual conversation.

Dynamic Prompt System

Seven intelligence layers assembled per response.

Every AI request is built fresh by the Dynamic Prompt System using seven layers assembled in strict order: (1) Identity Core — personality, expertise boundaries, species knowledge; (2) Response Intelligence — diagnostic logic, accuracy guardrails, anti-repetition; (3) Emergency Protocol — three severity tiers with context-aware detection; (4) Engagement Rules — follow-up question generation, tone control, product recommendations; (5) Conversation Protocol — greeting behavior, memory usage rules ("use silently, never announce"); (6) Pet Context — name, species, breed, age, weight, sex, location, seasonal awareness; (7) Memory Context — top memories by importance, silently integrated.

Why it matters: Static prompts waste tokens on irrelevant context and miss the slices that matter for the current question. The system detects query type (health vs casual via 30+ medical keywords), adjusts the token budget accordingly, and ensures every user gets the right context at the right cost — scaling model quality without scaling cost linearly.

BossRadar Context Engine

Priority-scored real-time health injection — allergies pinned first for safety.

BossRadar is the in-house context engine that surgically injects each pet’s medical reality into every Claude prompt. Every item on the pet’s record is priority-scored: chronic conditions score 100 (always injected), diagnosed conditions 50, new concerns 25, behavior items 10 — with modifiers for severity (+5), recency (+3 if under 30 days), and medication association (+2). The output is constrained to a strict 320–670 token budget per request, with guaranteed inclusion of ALL chronic conditions, backfill to six ongoing items, the three most recent resolved items, ALL active medications and ALL allergies.

Why it matters: Allergies are injected first — a deliberate safety design using the primacy effect so the AI reads allergy data before anything else, even under tight budgets. Generic AI hallucinates because it has no idea which animal you’re talking about. BossRadar gives Claude the exact medical context for this pet at this moment, before a single token of the response is generated.

Memory Architecture

Three-phase learning system — extraction, curation, and a living profile that evolves with every conversation.

Phase 1 — Extraction. At messages 5, 10, 15 and 19 of every conversation, Claude Sonnet extracts up to five owner-stated facts categorized across eight memory categories (training, behavior, nutrition, environment, social, exercise, bonding, health history). Message 10 runs a dual-duty call — extracting memories AND generating a 200–300 word conversation summary. The AI can emit memory flags to correct contradictions via fuzzy-match handling; importance decays automatically for low-value facts.

Phase 2 — Curation. A five-task pipeline powered by Claude Haiku runs three times a week on Vercel Cron: health keyword purge → importance decay → health history summary → merge & deduplication → curated profile generation. Phase 3 — Profile. The output is a single curated profile document covering personality, daily life, nutrition, training, social, exercise, owner bond and health history — sized at 400 tokens on Boss Pro and 600 tokens on Boss Super, injected silently as context into every future conversation. The AI uses this knowledge without ever announcing “I remember.”

Why it matters: Most chatbots have zero memory between sessions. BossPaws AI extracts, compacts, merges and curates long-term memory into a living profile that evolves with every conversation — so the AI gets smarter about your pet the longer you use it.

Video Analysis

The crown jewel — 30 seconds of video, dual-pipeline veterinary observation.

No pet health app in the world analyzes video. BossPaws uses a dual-pipeline architecture: Google Gemini Flash produces an exhaustive veterinary observation report covering animal identification, body condition (coat, skin, eyes, ears, nose), movement and gait (limping, balance, coordination), breathing (rate, effort, chest rise and fall), posture (stance, weight-bearing, body tension), behavior (alertness, pain indicators, energy) and environment (context, hazards). Gemini’s instruction is explicit: “Be exhaustive — describe everything you observe, never diagnose.” Then Claude Sonnet generates a clinical-grade assessment adapted by purpose — health-focused, training-focused, behavior-focused, or appearance-focused.

Why it matters: Thirty seconds of your pet walking, playing, or breathing gives the AI clinical-grade observations that would normally require a vet visit. The Gemini observation is cached in the database — reusable for multiple analyses without re-processing. Cost-optimized: Gemini handles heavy multimodal lifting, Sonnet handles clinical reasoning. Limits: 30 seconds maximum, 50 MB maximum, Boss Pro 6/week, Boss Super 12/week.

Multimodal Document Intelligence

Photos, PDFs, vet records, lab panels — analyzed with full health context. No Gemini — Sonnet Vision throughout.

Pet owners upload everything: phone photos of paper lab results, blurry medication labels, scanned x-rays, prescription PDFs, vet discharge instructions. Multimodal Document Intelligence splits into two clean paths, neither of which involves Gemini. Images: Claude Sonnet Vision first classifies each upload into one of four content types — pet photo, vet document, medical image, or non-pet — with confidence scoring that rejects non-pet content and low-confidence detections. Type-specific analysis follows: body condition for pet photos, lab value interpretation for vet documents, severity assessment for medical images. PDFs: CloudConvert (Poppler engine) renders up to six pages into 1024×1024 pixel JPEGs at 300 DPI; Sonnet Vision then reads all pages together, maintaining context across the full document. Text documents are processed directly (UTF-8, UTF-16, Latin-1).

Why it matters: Every analysis receives the pet’s full profile plus BossRadar health context — conditions, medications, allergies. Upload a blood panel and the AI reads it, flags abnormals, compares to your pet’s known health baseline, and suggests questions for your vet — all while knowing the pet’s breed, age, medications and allergies. Critical lab values trigger emergency alerts automatically.