SmarterDoc AI

SmarterDoc AI
Frontend Flow
Backend Flow

Project Story

Inspiration

Healthcare data today is fragmented and overwhelmingly complex. Patients lack a unified way to find a compatible doctor that checks off all their specific requirements. We still have to google for a doctor’s patient testimonials and track record. SmarterDoc AI was inspired by a simple idea: finding the right doctor should be seamless, not manual.

Our goal was to combine grounded searched data with AI-driven personalization and auto-appointments to make healthcare navigation easier and more efficient.

What It Does

SmarterDoc AI is an intelligent healthcare search and recommendation platform that connects patients and physicians through our multi-agent flow. The platform allows users to:

Search for doctors by specialty, insurance, and location.
View verified doctor profiles from the NPI Registry and PubMed.
Visualize nearby providers on an interactive Google Map.
Receive AI-generated recommendations that base decisions on provider profile and query similarity with dynamic weighted ranking.
Make automated phone-call appointments through our telephony agent.

Our application is built on data powered by Fivetran and the GCP ecosystem.

How We Built It

Layer	Description	Technologies
Frontend	Web interface for search, filtering, and booking	Next.js 14, Tailwind CSS, TypeScript
Backend	API layer for ranking, retrieval, and data analysis	FastAPI (Python 3.11+), Docker
Infrastructure	Deployment and CI/CD	Google Cloud Run, GitHub Actions
Data & AI	Data pipelines, analytics, and AI recommendations	Fivetran, BigQuery, Vertex AI, Cloud Storage

Data Integration

Built three custom connectors using the Fivetran Connector SDK to extract provider data and enrichment with Gemini grounded search.
Loaded the data into BigQuery and deployd vector embeddings to Vector Search Endpoint, processed with Vertex AI.
Enabled scheduled flow and checkpoints to update new provider's information.

Search and AI Layer

Vertex AI embeddings for vector similarity search on provider profile and user query in the first filter layer.
Agent-assigned dynamic contextual weights to provider features based on user's query in the ranking calculation; the feature scores are then used along with provider composite profiles for the agent to make the final selection.

Cloud Infrastructure

Deployed both frontend and backend on Google Cloud Run using Docker containers.
Managed continuous deployment with GitHub Actions.
Used GCP Secret Manager for configuration security and scalability.

Challenges We Faced

Merging data from multiple public sources with inconsistent schema and building the Fivetran connectors with sophisticated data processing pipelines.
Balancing AI explainability and performance in doctor ranking models and optimizing deterministic output for similar queries.
Reducing RAG agent search latency on provider selection.
Optimizing container startup times for Cloud Run cold starts.

What We Learned

Building a data-driven AI application requires equal attention to infrastructure, data quality, and user trust.
Transparent recommendations are essential; AI should be explainable and the results should be backed.
Fivetran pipelines can significantly reduce ETL complexity and work seamlessly with GCP ecosystem like Google BigQuery and Vertex AI.
Fivetran has an agent script where you can post on Gemini Gem and provide it a knowledge base to quickly spin up a prototype connector using the Fivetran SDK (workability not guaranteed).
Vertex AI provides powerful and scalable hybrid vector search capabilities.

What’s Next

Integrate Gemini API for conversational healthcare recommendations.
Expand regional data sources for better coverage and cost prediction.
Launch a public dashboard for real-time healthcare analytics (using Fivetran to read in user appointment data).
Collaborate with research partners to evaluate model fairness and bias reduction.

Built With

ai-agents
bigquery
docker
fastapi
fivetran
gcp
gemini
github
googlemapapi
llm
multi-agents
next.js
npiregistryapi
pubmedapi
rag
tailwind
typescript
vertexai

Submitted to

AI Accelerate: Unlocking New Frontiers

Created by

I worked on backend API endpoints and built the custom Fivetran connectors and data enrichment pipeline for provider profiles. Profile embedding vectorization and vector indexing. I also implemented the recommendation agent and ranker algorithm for the profile feature scores.

Yian Chen
I developed an intelligent medical appointment system that integrates speech-to-text and AI voice calling capabilities. The system uses Google Cloud Speech-to-Text for real-time audio transcription and implements automated phone calls to doctors through Twilio and Vertex AI Live API for appointment scheduling. The platform also includes doctor search, intelligent ranking, and AI chat functionality to provide patients with a comprehensive medical appointment service.

Zirui Wen
I led the development of the entire frontend using Next.js, Tailwind CSS, and TypeScript, building the full user experience for doctor search, AI recommendations, and appointment booking. I also implemented backend CI/CD pipelines with GitHub Actions and deployed both the frontend and backend on Google Cloud Run using Docker and GCP Secret Manager. In addition, I participated in the Vertex AI and Fivetran integration discussions, helping design how AI-driven ranking and data ingestion flow through BigQuery and Vertex AI. I also handled video editing and production for the project demo.

Hogan Lin
I’m responsible for the entire project’s interface design, including structuring the system’s information architecture and user flow, defining the overall visual style, and building modular components to ensure design consistency and improve development efficiency.

Wan Chen Wu