Triage Flow Review
A Next.js + FastAPI research prototype for multimodal medical triage using ResNet18 image classification, FAISS RAG, and Google ADK agents.
Each review covers deployability, value versus commercial alternatives, and privacy model. Tools that can run locally were started and exercised; mobile or backend-dependent tools were assessed from published builds, source code, and deploy guides. Ratings reflect what we were able to verify.
Triage Flow — Multimodal AI Medical Triage System
A Next.js + FastAPI research prototype for multimodal medical triage using ResNet18 image classification, FAISS RAG, and Google ADK agents. Last updated: 2026-06-21.
One-sentence verdict: An ambitious research-oriented triage prototype that is harder to deploy than simpler chatbots because it requires Supabase, Gemini, Heygen, and OpenAI keys.
What the System Is
Triage Flow is a multimodal medical triage system. It is built as:
- Next.js + Tailwind CSS + shadcn/ui frontend.
- FastAPI backend in Python.
- TensorFlow ResNet18 fine-tuned on DermNet for skin-condition image classification.
- FAISS vector store with medical PDF RAG.
- Google Agent Development Kit (ADK) with two agents: information retrieval/summarization and severity assessment.
- Supabase for authentication, database, and storage.
- Gemini API, Heygen API, OpenAI API, and Resend API for voice, video, and email features.
The deploy guide lists the project as documented-only because it requires external Supabase and API keys to run fully.
| Key data | |
|---|---|
| Category | Symptom Checker / Triage |
| Language | TypeScript / Next.js + Python / FastAPI |
| License | Unknown |
| Self-hosted | Partial (requires Supabase cloud) |
| AI | Google Gemini, ResNet18, Google ADK |
| Database | Supabase (PostgreSQL) |
| Deployment | Local dev servers + Supabase |
How to Install and Deploy
The deploy guide documents local setup but notes that Supabase and several API keys are required.
# 1. Configure Supabase and run schema.sql from frontend/lib/supabase/migrations
# 2. Start backend
cd /data2/docker/going-global/repos/triage-flow/backend
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
uvicorn api.triage_endpoint:app --host 0.0.0.0 --port 8000
# 3. Start frontend
cd /data2/docker/going-global/repos/triage-flow/frontend
npm install
npm run dev
Required environment variables include SUPABASE_URL, SUPABASE_KEY, GOOGLE_API_KEY, MODEL_GDRIVE_ID, NEXT_PUBLIC_SUPABASE_URL, NEXT_PUBLIC_SUPABASE_ANON_KEY, HEYGEN_API_KEY, OPENAI_API_KEY, and RESEND_API_KEY.
How to Test
The documented test flow is:
- Create and configure a Supabase project.
- Run
schema.sqlin the Supabase SQL editor. - Start the backend on port 8000 and frontend on port 3000.
- Open
http://localhost:3000, register, and log in. - Submit a triage request with symptom text and optionally an image of a skin condition.
- Review the severity score, reasoning, and recommendations returned by the ADK agents.
- Optionally test the voice assistant and video consultation features if the required API keys are configured.
Privacy & Compliance
Triage Flow is not HIPAA compliant. It relies on multiple third-party cloud services (Supabase, Google Gemini, Heygen, OpenAI, Resend) and sends symptom text and images to those services. Do not upload real patient data or PHI.
Triage Flow vs Commercial Symptom Checkers
| Dimension | Triage Flow | Commercial Triage (e.g., Ada Health, K Health, Buoy Health) |
|---|---|---|
| Cost | Free / open-source (API costs apply) | Subscription or bundled with care |
| Data location | Supabase + multiple AI clouds | Vendor cloud |
| Multimodal input | Text + skin images | Varies |
| Clinical validation | Research prototype | Clinically validated, regulated in many cases |
| Setup effort | High: Supabase, model download, multiple API keys | Low: app install |
| Voice/video avatar | Heygen + Gemini Live | Usually not included |
| Open source | Yes | No |
Who Should Use It
- Researchers and developers exploring multimodal triage with image classification and RAG.
- Teams evaluating Google ADK for medical-agent workflows.
- Builders who already have Supabase and AI provider accounts and want a research scaffold.
Who Shouldn't Use It
- Users who need a simple, quick-to-run symptom checker.
- Anyone who cannot or will not configure Supabase and multiple paid API keys.
- People seeking clinically validated triage for real medical decisions.
FAQ
Is Triage Flow a finished product?
No. The deploy guide marks it as documented-only because it requires external Supabase and API keys. It is a research prototype.
What external services are required?
Supabase for auth/database, Google Gemini for LLM, Heygen for video avatars, OpenAI for additional AI features, and Resend for email. A Google Drive model file is also downloaded for the ResNet classifier.
Can I run it without images?
Yes, text-only triage is supported, but the multimodal image classification feature requires the downloaded ResNet model and image upload.
Verdict
Triage Flow is an ambitious research prototype that combines image classification, RAG, and agentic AI for triage. The complexity of its external dependencies makes it harder to deploy than simpler chatbots, and it is not suitable for real clinical use without substantial validation and compliance work.
Ratings: Deployability 2/5 · Value vs Commercial 3/5 · Privacy Compliance 2/5