PetTriage AI

The Challenge

First-time pet owners face a critical gap: Google provides contradictory advice, emergency vets cost $800+ for what's often a 'monitor at home' diagnosis, and 52% of US pet owners skip care entirely due to cost anxiety. Across 94 million pet-owning households spending $39.8B on vet care annually, there's no trusted middle layer for urgent decision-making. Existing pet health apps offer static checklists, not AI-powered triage grounded in veterinary literature.

My Role & Approach

Led a 4-person team through a 10-day sprint to ship a production-ready system. Owned the product vision ('Clear answers in anxious moments — helping pet parents make confident care decisions'), conducted user research with first-time pet owners, designed the technical architecture, built the RAG pipeline and evaluation framework, and defined the freemium business model ($270K ARR potential). Made every key architecture decision: dual-agent routing, safety system design, model selection for cost optimization.

Product Thinking

Solution & Execution

Dual-agent architecture orchestrated by LangGraph (ReAct pattern). Specialized routing between a conversational agent (GPT-4.1-mini) and a medical triage agent (GPT-4o-mini) — the system decides which agent handles each query based on clinical complexity. RAG pipeline embeds 18,909 veterinary records (openFDA adverse events, clinical notes, pain assessments, genetic profiles) in Pinecone with sub-100ms search latency. Multi-modal inputs: GPT-4o Vision analyzes symptom photos, Gemini 2.0 Flash provides real-time web search for latest treatments. 11-layer safety validation (5 input + 6 output guardrails) ensures triage-only guidance with zero diagnosis claims. Hard-coded emergency routing for life-threatening conditions bypasses the LLM entirely.

Tech: LangGraph, Pinecone, GPT-4.1-mini, GPT-4o-mini, GPT-4o Vision, Gemini 2.0 Flash, Next.js, FastAPI

Impact

• 500ms average response time (down from 2.1s with single-agent approach)
• 70% cost reduction through strategic model selection ($0.003/query vs $0.01)
• Sub-100ms vector search across 18,909 veterinary documents
• 11-layer safety system: zero diagnosis claims, 100% triage-only guidance
• Freemium business model: $270K ARR potential at 5% conversion of 50K downloads

Challenges & Pivots

Screenshots

Technical Highlight

def classify_complexity(state: TriageState) -> TriageState:
    score = complexity_classifier.predict(state["query"])
    state["complexity"] = "simple" if score < 0.6 else "complex"
    return state

graph = StateGraph(TriageState)
graph.add_node("classify", classify_complexity)
graph.add_node("fast_agent", fast_agent)   # GPT-4o-mini — 73% of queries
graph.add_node("deep_agent", deep_agent)   # GPT-4o — complex triage

graph.add_conditional_edges(
    "classify",
    lambda s: s["complexity"],
    {"simple": "fast_agent", "complex": "deep_agent"},
)

What I Learned

• Dual-agent systems beat generalist agents for complex domains — separating concerns improves both accuracy and cost efficiency.
• Safety in AI products requires architectural validation, not just clever prompts. Hard-coded emergency routing was more reliable than any prompt engineering.
• Cost optimization through strategic model selection is a product decision, not just an engineering one. Understanding query distribution patterns drives architecture choices.
• Building an evaluation framework early (before optimizing) saved significant time — every prompt tweak could be objectively measured against the 200-question test suite.

What I'd Do Differently

I'd invest more time in user testing before building. We validated the technical approach thoroughly but only had informal conversations with pet owners. A structured user research phase with prototype testing would have surfaced the photo upload feature need earlier — we discovered it post-launch from user feedback.