Sanare-Health

Sanare transforms symptoms into evidence-backed, personalized health insights via multi-agent reasoning & real medical citations, helping users understand what’s happening & why, not just what to do.

Comment

What it does

🌱 About Sanare

We were inspired by the gap between quick symptom checkers and trustworthy medical reasoning. Most digital tools focus on what might be wrong, Sanare focuses on why. Derived from the Latin word for “to heal,” Sanare combines empathy, explainability, and real medical evidence to help users make confident, informed health decisions.

⚙️ How We Built It

Sanare is powered entirely by Google Cloud Platform, using a modular multi-agent architecture where each agent contributes a unique layer of reasoning:

  • Vertex AI (Gemini models): Hypothesis, Context, Research, Contradiction, Synthesis, and Alert Agents.
  • Cloud Firestore: Structured storage for user profiles, encounters, and agent outputs.
  • Cloud Storage (GCS): Saves large reasoning graphs, citation JSON, and model logs.
  • Firebase Authentication: Provides secure sign-in and session handling.

The frontend was developed using Next.js + TypeScript + TailwindCSS, hosted on Firebase Hosting with automatic SSL.

🧠 Sanare’s Multi-Agent System

Agent Purpose Input Output
🧩 Hypothesis Agent Converts user-entered symptoms and demographics into a structured, measurable clinical hypothesis. Raw user text (primary symptom, duration, severity). Normalized symptom + potential body system (e.g., Chest pain → Cardiovascular hypothesis).
🌍 Context Agent Enriches the hypothesis with personal context — age, sex, pregnancy, medical history, allergies, and medications. Hypothesis + user profile from Firestore. Personalized symptom framing (“Chest pain in a 45-year-old diabetic male on statins”).
📚 Research Agent Retrieves real-time evidence from trusted medical sources (PubMed, CDC, Mayo Clinic, NIH). Structured query from Context Agent. Key findings + citation JSON (PubMed IDs, article links).
⚖️ Contradiction Agent Performs risk analysis by identifying conflicting patterns or red-flag indicators that could invalidate the hypothesis. Evidence list + user inputs. Risk flags, contradictions, and alternative hypotheses.
🧮 Synthesis Agent (Future Work) Balances all findings to compute a confidence score, ranking possible conditions and recommended next steps. Hypothesis + Context + Research + Contradictions. Confidence (0–100%), top differential list, urgency level (self-care / routine / urgent / emergency).
🚨 Alert Agent (Future Work) Monitors for emergency criteria (e.g., severe chest pain, breathing difficulty, stroke signs) and triggers immediate safety guidance. Real-time user input stream. Emergency alert + action plan (“Call 911 / Visit ER”).

🧠 What We Learned

We learned how to design explainable AI workflows using Vertex AI, create a FHIR-inspired Firestore schema separating stable profile data from dynamic triage sessions, and use BigQuery for lightweight, privacy-safe analytics. Our biggest insight: transparency and evidence citation dramatically increase user trust in health AI.

🚧 Challenges

  • Aligning multi-agent reasoning with real clinical guidelines while keeping latency low.
  • Managing dependency conflicts between React, Next.js, and shadcn/ui.
  • Building a UI that is calm, inclusive, and fully accessible (WCAG AA).

💡 Takeaway

Sanare demonstrates how transparent, evidence-based AI health companions can be built responsibly with Google Cloud - merging reliability, reasoning, and empathy in one experience.

Built With

Share this project:

Updates