The problem
Veterinary care often loses visibility the moment a consultation ends. A clinic may provide a clear care plan, but once the owner goes home, follow-up becomes fragmented: check-ins are missed, observations arrive late or unstructured, and the veterinarian has to reconstruct what happened from messages and memory.
What we built
CaniBiz CareLoop Agent is an autonomous veterinary follow-up workflow that keeps the care plan moving after the consultation.
Instead of acting like another chatbot, CareLoop owns the follow-up workflow:
- structures a veterinary care plan into actionable follow-up tasks,
- maintains persistent workflow state,
- detects missed check-ins,
- autonomously requests the next owner update,
- records structured observations,
- runs a deterministic safety gate for predefined escalation signals,
- and generates a concise VetBrief for professional review.
The consultation ends, but care doesn’t.
Why it is agentic
CareLoop does not wait for every next prompt. It tracks workflow state and can determine when a scheduled follow-up has been missed, then generate the next required action.
The system separates responsibilities deliberately:
- Gemini interprets and structures natural-language care instructions and owner updates.
- Google ADK orchestrates the agent workflow.
- Persistent state preserves tasks, observations and workflow progress across steps.
- Deterministic safety policy handles high-risk escalation independently from the LLM.
- VetBrief compresses the longitudinal follow-up into a professional handoff.
Golden demo: Milo
Our demonstration follows Milo, a 6-year-old Labrador undergoing a five-day diet transition with daily stool monitoring.
The care plan becomes five structured follow-up tasks. Milo progresses normally, then a scheduled Day-3 check-in is missed. CareLoop detects the overdue task and requests an owner update without requiring the owner to initiate a new conversation first.
The owner later reports that Milo ate normally but his stool was softer. The observation is persisted and the workflow continues.
On Day 4, the owner reports: “He vomited twice and seems unusually tired.”
The system normalizes the observation, while the deterministic safety layer identifies repeated vomiting and marked lethargy and escalates the case for professional review. The LLM cannot downgrade that safety result and does not diagnose Milo.
CareLoop then generates a VetBrief so the veterinary professional can review the relevant timeline without reading an entire conversation.
Safety architecture
CareLoop is not an AI veterinarian and does not attempt autonomous diagnosis.
A core design principle is:
Gemini interprets. CareLoop orchestrates. Deterministic policy protects. Veterinarian remains the authority.
High-risk safety rules are evaluated outside the model, and professional review is required when escalation criteria are met.
Google technologies
The project is designed around the required Google stack:
- Gemini 3.5+ for structured interpretation and language tasks
- Google Agent Development Kit (ADK) for agent workflow orchestration
- Google Cloud Run for deployed application infrastructure
- Firestore for persistent workflow state
The exact model/runtime configuration used by the final build is documented in the repository and demonstrated in the project video.
Features
- Natural-language veterinary plan → structured follow-up tasks
- Stateful multi-day care workflow
- Autonomous missed-follow-up detection
- Structured owner observation capture
- Deterministic safety escalation
- Professional-review requirement for high-risk cases
- VetBrief generation
- Reproducible workflow demo
- Google Cloud deployment proof
What we learned
The most important lesson was that autonomy in a care workflow should not mean giving an AI more clinical authority.
A stronger architecture separates interpretation, orchestration and safety. The LLM is useful for turning messy human language into structured information, while deterministic policy is better suited to non-negotiable safety boundaries.
We also found that persistent workflow state matters as much as model capability. An agent only becomes genuinely useful when it can remember what task is due, what was missed, what the owner actually reported, and what needs to happen next.
Current scope
This hackathon project uses synthetic demonstration data and a standalone workflow architecture. It does not expose production CAIOS data and is not presented as diagnostic or clinical decision software.
Future direction
The same workflow can later connect, through a controlled context interface, to a longitudinal companion-health system so that professional follow-up can benefit from richer historical context without coupling the hackathon application directly to production health data.
Built With
- fastapi
- firestore
- gemini-3.5+
- google-adk
- google-cloud-run
- python