mongo-db-atlas-search
mongo-db-csv-records
mongodb-multimodal-embeddings
google cloud storage articke page bucket
ui
diagram

Search4Cure.AI-Diabetes

Inspiration

Diabetes is a chronic condition affecting millions globally, with a growing body of medical research being published every day. However, accessing and making sense of the vast amounts of research literature remains a significant challenge for both clinicians and researchers. We were inspired by the need for a smart, accessible AI-powered assistant that could help accelerate scientific discovery and decision-making in diabetes care and research. The idea was to combine multiple specialized agents, each with a distinct role, to collaborate like a research team—one that never tires and is always up to date.

What it does

Search4Cure.AI-Diabetes is a multi-agent research assistant designed to help users explore, synthesize, and extract insights from diabetes-related medical literature. The system accepts user queries and uses a team of LLM-powered sub-agents to handle document retrieval from embedded arxiv articles, summarization, csv file retrieval, multimodal reasoning (e.g., interpreting charts and figures), and insight generation. The agents collaborate via a central controller to provide a coherent and thorough response, along with a user-friendly interface that supports both natural language questions and structured queries.

How I built it

We developed the project using the ADK (Agent Developer Kit) framework provided by Google, which enabled us to compose and orchestrate multiple agents with ease. Here's a high-level overview of the tech stack and approach:

Python 3.12 as the core development language.
Google ADK to manage and deploy cloud-native LLM agents.
LangChain and FAISS for document indexing, retrieval, and chunk-level semantic search.
** Gemini** for powering the reasoning and summarization capabilities of the agents.
Google Cloud Run for deploying a scalable backend service.
Streamlit and ADK UI components for building an interactive frontend experience.
Docker for containerization and cross-environment compatibility.

Each sub-agent was designed with specific skills.

Challenges I ran into

Multimodal analysis: Extracting and reasoning over images (graphs, tables) required integrating OCR and vision-language models, which added layers of complexity.

Accomplishments that I'm proud of

Successfully deployed a production-ready multi-agent system on Google Cloud.
Enabled seamless natural language interaction with scientific content.
Integrated document-level and visual (multimodal) analysis into a unified agent workflow.
Designed a modular backend that allows easy extension to other medical domains beyond diabetes.

What I learned

The power and limitations of large language models in specialized domains like medicine.
How to design modular, collaborative LLM agents using Google’s ADK and best practices.
Cloud-native deployment strategies for AI agents, including Dockerization, service exposure, and permission scoping.
How to integrate RAG (Retrieval-Augmented Generation) and multimodal tools into a cohesive research assistant.

What's next for Search4Cure.AI-Diabetes

Broaden disease scope: Extend the system to cover other chronic diseases like cancer, cardiovascular disease, and Alzheimer’s.
Fine-tune on medical corpora: Train custom LLMs or adapters on medical literature for even deeper domain understanding.
Multilingual support: Add capabilities to process and summarize literature in languages other than English.
Research collaboration assistant: Allow the assistant to help co-author scientific reports or grant proposals.
Patient-facing features: Build a simplified frontend for patients seeking easy-to-understand summaries of the latest research.

Search4Cure.AI-Diabetes is just the beginning of building intelligent agents to accelerate medical understanding and empower both researchers and patients.