• 

EvidencePilot

EvidencePilot is an autonomous, evidence-focused clinical research agent built with Gemini and the Google GenAI SDK.

An AI-powered evidence research agent that turns complex clinical questions into a structured research workflow.**

EvidencePilot helps transform an open-ended clinical research question into a research plan, retrieves relevant literature, organizes available evidence, and identifies evidence gaps and uncertainties.

Inspiration

Clinical research can require researchers to move through many sources, extract relevant findings, compare evidence, and identify unanswered questions. For complex questions, this process can become repetitive and difficult to organize.

With a medical background and a growing interest in AI and research, I wanted to explore a practical question: Can an AI agent take on the repetitive parts of an evidence-research workflow while keeping the researcher in control?

That idea became EvidencePilot.

What We Are Building☆☆

EvidencePilot is an AI powered clinical research agent designed to turn complex clinical research questions into structured, evidence-focused research workflows.

Instead of simply answering a question like a chatbot, EvidencePilot is designed to perform a multi-step workflow: understand the research question, plan the investigation, gather relevant evidence, extract important findings, evaluate the evidence, identify gaps, and synthesize the results.

The goal is to give researchers a useful starting point for deeper investigation while keeping sources and uncertainty visible.

How It Works☆☆

A user provides a complex clinical research question.

EvidencePilot then breaks the question into research tasks and coordinates an agentic workflow to investigate those tasks. The system collects relevant evidence, extracts useful information, organizes findings, and produces a structured research brief.

The final output is designed to make it easier for a researcher to see:☆☆

• What evidence was found
• Which sources support the findings
• Where evidence is limited or conflicting
• What important questions remain unanswered

Why An Agent?☆☆

A conventional chatbot can generate an answer in one step. EvidencePilot is designed around the idea that research is a workflow, not a single question.

The agent therefore needs to plan, act, evaluate intermediate results, and synthesize information before producing its final brief.

This makes the project a practical exploration of how agentic AI can support knowledge-intensive work.

Technology☆☆

EvidencePilot was built with Python and FastAPI, using the Google GenAI SDK and Gemini for AI capabilities, PubMed/NCBI for evidence retrieval, and Docker for containerization.

The project was also deployed and tested with Google Cloud Run. ,Render.

What I learned

A major focus of the project is reliability.

AI-generated research assistance is only useful when users can understand where information came from and where uncertainty remains. We therefore want EvidencePilot to make source traceability and limitations visible rather than presenting generated conclusions as unquestionable facts. I am also exploring how to balance autonomous task execution with appropriate human review. ▪︎Building EvidencePilot helped me learn how to combine AI reasoning with external evidence retrieval, structure an agentic research workflow, build an API with FastAPI, containerize an application with Docker, and deploy a working AI application.

The project also taught me an important practical lesson: building an AI agent is not only about the model—it is about designing a reliable workflow around the model, evidence sources, uncertainty, and reproducible execution.

Agentic Workflow

EvidencePilot is designed as a workflow rather than a single prompt-response system. Each stage has a specific role:

• Planner — breaks the clinical question into research objectives.
• Retriever — searches relevant PubMed literature.
• Evidence Organizer — structures retrieved evidence.
• Gap Analyzer — identifies missing evidence and uncertainties.
• Synthesis Layer — prepares the evidence-focused research output.

This separation makes the workflow easier to inspect, reproduce, and extend.

Challenges

The main challenge is building something that is genuinely agentic rather than simply placing a chat interface around a language model.

Another challenge is maintaining evidence traceability while asking an AI system to perform multiple research steps. The system must distinguish between retrieved evidence, extracted information, and AI-generated synthesis.

Testing

EvidencePilot was tested using clinical research questions through its deployed FastAPI Swagger interface. The API successfully returned HTTP 200 responses and produced structured research plans and evidence-gap outputs.

Example test question:

What is the cardiovascular benefit of GLP-1 receptor agonists in adults with type 2 diabetes?

The system generated research objectives and explicitly reported when additional evidence retrieval was required.

What's Next

I plan to test EvidencePilot on representative clinical research questions, improve its research planning and evidence evaluation workflow, and refine the final user experience.

The longer-term goal is to explore whether agentic AI can reduce repetitive research work while helping researchers spend more time on interpretation, critical thinking, and new questions.

Built With

• agenticai
• aiagents
• artificialintelligence
• gemini
• generativeai
• google-cloud
• google-research-search
• healthcare
• machine-learning
• medicalresearch