Guardian — AI-Powered Emergency Companion
Hackathon Project · 2025
Overview
Guardian is an AI-powered emergency companion designed for people living alone. The core problem it solves is simple but critical: when a medical emergency happens, every minute matters. Guardian bridges the gap between an incident and receiving help by continuously monitoring health signals, detecting potential emergencies, communicating with the user, notifying caregivers, and escalating situations when immediate action is needed.
What made this project especially compelling was the decision to run entirely on-device. Rather than relying on cloud services, Guardian performs all its reasoning locally — keeping sensitive health data private while delivering fast, reliable responses precisely when they matter most.
Architecture
The agent system was built with a multi-agent architecture using LangGraph as the orchestration framework. The system consists of one orchestrator agent and six specialized sub-agents, each responsible for a distinct class of patient needs:
- Health signal monitoring and anomaly detection
- Emergency escalation and caregiver notification
- Direct calls to emergency contacts via Twilio
- Database navigation for patient records and medical history
- Online class and appointment scheduling
- General patient assistance and communication
Each sub-agent had access to its own curated set of tools drawn from the project's shared tool registry, keeping responsibilities clearly separated while allowing the orchestrator to route tasks intelligently.
Observability
LangSmith was used for full agent observability — tracing every step of each agent's reasoning, tool calls, and decisions. This gave the team real-time visibility into agent behavior during development and made it significantly easier to debug edge cases in emergency scenarios where correctness is non-negotiable.
My Role
I was responsible for designing and building the entire agent system — from the orchestrator's routing logic to the per-agent tool integration. This included:
- Designing the orchestrator-to-subagent delegation strategy in LangGraph
- Integrating Twilio for real-time emergency contact calls
- Wiring up database read/write tools across relevant sub-agents
- Setting up LangSmith tracing for end-to-end observability
Tech Stack
LangGraph · LangSmith · Twilio · Python · On-device LLM inference