CareMesh

Inspiration

We were inspired by a simple gap in care: the warning signs usually show up early, but they are scattered. Sleep drops, stress rises, routines break, activity falls off, and check-ins start sounding heavier. Most tools either track data without context or give generic advice without understanding who the person is.

CareMesh was built to bridge that gap for people who are especially easy to overlook in stressful seasons, including students, caregivers, older adults, and users who need lower-effort, more accessible support. We wanted a system that could catch strain early and turn it into calm, practical next steps before a hard day becomes a crisis.

What it does

CareMesh is a companion platform that combines wearable data, manual check-ins, and persona-aware care coordination. A user can sign in, complete onboarding, connect Garmin, submit a manual check-in, or run a seeded scenario. The platform then reads recent signals, interprets them in context, and creates a support plan with meal guidance, wellness actions, empathy-first messaging, and follow-up when risk is elevated.

It also includes a health dashboard, recipe parsing and storage, recommended recipes, a coordinator dashboard for open cases, and a trace view that shows how the support plan was created step by step.

How we built it

We built the frontend in apps/web using React, TypeScript, React Router, TanStack Query, Tailwind, and shadcn/ui. The app includes onboarding, member and coordinator dashboards, a health dashboard with Garmin connection and sync controls, recipe pages, a scenario runner, and an agent trace view.

We built the backend in apps/api using FastAPI, PostgreSQL, SQLAlchemy async, and Alembic. Clerk handles identity, while the backend links each Clerk user to an internal user record and stores profiles, health data, events, interventions, cases, notifications, recipes, meal plan slots, and audit logs.

For the intelligence layer, we built a Google ADK-aligned multi-agent pipeline under services/agents, services/tools, and services/remote_specialists. The coordinator pulls a profile and recent signals, runs signal interpretation, risk stratification, and intervention planning in parallel, routes to persona-aware specialists, generates empathy-first language, runs a validation loop, and persists the final intervention and case actions.

Challenges we ran into

One of the biggest challenges was integration. The frontend and backend were developed in parallel, so we had to preserve the frontend experience while replacing mock data with real API contracts, real authentication, and real backend state.

Authentication was another major challenge. We moved the app to Clerk-only auth, linked Clerk identities to internal users, and had to debug backend token verification, onboarding state, local development origins, and session sync between frontend and backend.

We also ran into data-shape and workflow challenges. Garmin sync, health metrics, recipes, interventions, and agent traces all had to feel like one product instead of separate features. Getting those systems to share the same user identity, data ownership, and execution flow took a lot of careful backend design.

Accomplishments that we're proud of

We are proud that CareMesh is not just a UI prototype. It is a working full-stack platform with real authentication, real persistence, real agent orchestration, and real traceability.

We are also proud of the multi-agent flow. The system does more than generate one answer. It interprets signals, stratifies risk, plans interventions, routes to persona-aware specialists, validates the plan, and records the process in a way that judges and coordinators can actually inspect.

We are especially proud that the project stays grounded in care. The recommendations are designed to be realistic, supportive, and persona-aware, not just technically impressive.

What we learned

We learned that good care technology needs strong foundations before it needs more model complexity. Identity, data contracts, persistence, safety checks, and traceability matter just as much as the intelligence layer.

We also learned that empathy works best when it is structured. A supportive message only becomes useful when it is backed by real signals, a clear understanding of the user, and practical next steps that someone can actually follow.

Finally, we learned that multi-agent systems are most valuable when they are tied to product workflows. The agents are strongest when they are part of a real platform with dashboards, cases, recipes, health views, and follow-up, not just a model demo.

What's next for CareMesh

The next step is to make the loop between Garmin data, health history, and agent decisions even tighter by projecting synced wearable data directly into the signal pipeline the agents use. That will make the recommendations feel more real-time and less demo-driven.

We also want to strengthen recipe recommendations, expand template and meal-planning support, and improve how interventions map to concrete actions a user can save and follow.

Beyond that, we want to make the model layer provider-neutral, improve deployment readiness, and deepen the coordinator workflow so CareMesh can move from a strong hackathon system into something closer to a production-ready care support platform.