BioqueryAI

Inspiration

In the biomedical world, knowledge evolves faster than humans can read. Every day, thousands of new research papers are added to PubMed, making it overwhelming for scientists and students to stay up to date. During my own research journey, I often struggled to find relevant papers quickly — manually scanning abstracts, comparing results, and summarizing findings took hours. This sparked the idea for BioQueryAI — an AI-powered assistant that could think like a researcher and summarize biomedical insights instantly.

What it does

BioQueryAI is an intelligent research assistant that connects directly to PubMed’s live API to retrieve and summarize the most relevant biomedical literature. It uses AI models to: Dynamically fetch the latest publications related to a query. Extract and rank papers using semantic similarity. Generate concise, human-like summaries with citations and context. Researchers can ask natural-language questions like: “What are the recent advances in AI-driven cancer drug discovery?” Within seconds, BioQueryAI delivers a synthesized summary built on real PubMed data.

How we built it

The system is designed as a retrieval-augmented generation (RAG) pipeline with three main layers:

Retrieval Layer — Fetches real-time data from the PubMed API.

Embedding Layer — Uses NVIDIA Retrieval Embedding NIM to encode and rank results semantically.

Reasoning Layer — Summarizes findings using Llama-3 NIM for natural language generation.

The frontend is built with Streamlit, and the entire pipeline is containerized with Docker and deployable on AWS SageMaker for scalable inference.

Challenges we ran into

Import path issues during modular app development in Streamlit.

Managing real-time API retrievals without exceeding PubMed’s rate limits.

Handling unstructured biomedical abstracts and ensuring that AI-generated summaries remained factually accurate.

Integrating multiple NVIDIA NIM endpoints securely using .env configuration.

Each challenge strengthened the system’s reliability and deepened my understanding of AI-powered retrieval systems.

Accomplishments that we're proud of

Built a fully functional live biomedical assistant capable of summarizing real PubMed data.

Successfully integrated NVIDIA NIM microservices for embedding and inference.

Deployed a Dockerized Streamlit application that runs seamlessly on both local and cloud environments.

Designed a clean, research-friendly UI that transforms literature review into an interactive experience.

What we learned

How to implement retrieval-augmented generation (RAG) architectures from scratch.

The power of semantic embeddings for improving search relevance.

How to manage real-time data pipelines using APIs and large language models (LLMs).

Practical insights into AI ethics, ensuring summaries remain scientifically accurate and transparent.

What's next for BioqueryAI

Add support for PDF uploads and local paper analysis.

Enable visual dashboards for publication trends and keyword analytics.

Integrate citation graph visualizations to explore relationships between studies.

Expand the model to handle multimodal biomedical data — including figures, tables, and datasets.

Ultimately, BioQueryAI aims to become a personal AI research collaborator, accelerating discovery across biology, healthcare, and life sciences.

Built With

• ai
• amazon-ec2
• api
• containerization-docker
• data
• deployment-aws
• embeddings
• embeddings-sentencetransformers
• layer-tools-/-libraries-frontend-streamlit-backend-python
• models-llama
• openai
• sagemaker
• source-pubmed