MIRA

MIRA helps care teams move from reactive monitoring to focused, action-based care by turning patient engagement data into prioritized workflows and AI-assisted follow-ups.

Comment

Inspiration

The problem isn’t medicine. It’s operations. Healthcare workers spend up to 45 minutes every morning—spreadsheets, call logs, and manual tracking—to monitor patient engagement. This creates cognitive overload for care coordinators and by the time they identify a disengaged patient, the critical window for intervention has often closed.

We wanted to explore how AI could act like an always-on operations assistant that continuously reviews the entire patient population and highlights who needs attention first, reducing burnout while improving follow-up consistency.

What it does

MIRA is an AI care operations assistant that monitors patient engagement and generates a prioritized daily care queue for health teams.

Instead of a passive dashboard, MIRA acts as an active member of the team:

  • Automates "Morning Rounds": At 6:00 AM, MIRA audits the entire patient population, detecting engagement drops using time-based heuristics.
  • Intelligent Triage: It generates a finite, prioritized "Daily Care Queue," flagging urgent cases (e.g., "Missed 2nd Refill") vs. routine ones.
  • Autonomous Outreach: MIRA utilizes a Voice Agent to call patients in real-time, negotiating appointment rescheduling and logging the outcome directly back to the system without human input.
  • Safety-First Reasoning: Every AI decision is vetted by a "Critic" agent to ensure clinical safety before it reaches the patient or the dashboard.

How we built it

MIRA uses the Gemini API as the reasoning engine behind a multi-agent architecture.

We implemented a Supervisor–Worker–Critic system:

  • The Supervisor: Acts as the orchestrator, delegating tasks to specialist agents.
  • The Specialists: "Worker" agents analyze specific domains (e.g., Viral Load Trends, Social Determinants) using a Semantic Bucketing prompting strategy to extract structured data from unstructured clinical notes.
  • The Critic (Our Secret Sauce): A dedicated adversarial agent (the "Quality Doctor") that strictly reviews every proposed action against clinical guidelines. If a recommendation is unsafe, the Critic blocks it. This solves the "Hallucination" problem in healthcare.

The Voice (Interaction Layer): We used Gemini 2.5 Flash (via LiveKit) to power the real-time voice agent. This allows for sub-second latency, enabling the AI to have natural, empathetic conversations with patients about rescheduling and barriers to care.

The Body (Infrastructure): The backend is powered by FastAPI, SQLAlchemy, and PostgreSQL. We use Celery and Redis to handle the asynchronous "Morning Rounds" workflow, ensuring the heavy lifting happens in the background while the frontend remains responsive.

Challenges we ran into

  • Designing AI recommendations that are helpful but non-clinical
  • The "Hallucination" Risk: In healthcare, a wrong guess is dangerous. We solved this by architecting the Critic Agent—a distinct architectural layer solely responsible for rejecting unsafe AI outputs.
  • Latency vs. Empathy: Getting the voice agent to respond instantly while still sounding empathetic was difficult. Integrating Gemini 2.5 Flash Native Audio with LiveKit was the breakthrough that made the conversation feel human.
  • Making AI outputs concise enough for busy care workers

This led us to introduce the Critic agent as an adversarial safety layer.

Accomplishments that we're proud of

  • Architecting the "Critic": We didn't just build a chatbot; we built a self-correcting safety system.
  • Built a working end-to-end AI care operations prototype
  • Implemented a multi-agent safety architecture
  • Created an automated daily prioritization workflow
  • Demonstrated real task execution through the voice rescheduling agent
  • Designed the system specifically to reduce cognitive load, not just predict risk

What we learned

AI delivers the most value in care operations when it helps teams decide and act faster, not when it only produces predictions.

We also learned that trust requires:

  • Clear reasoning
  • Explainable flags
  • Human-in-the-loop workflows

What's next for MIRA

Next steps include:

  • We are preparing to pilot MIRA with a partner NGO to measure the real-world reduction in care worker burnout.
  • Expanding specialist agents to additional care programs
  • Integrating directly with existing EHR systems
  • Adding explainability views for why cases are prioritized

Built With

Share this project:

Updates

Submission history