NebulaPrime AI - Unlocking Next Gen Clinical Intelligence
NebulaPrime AI - Architecture Diagram

Project Story: NebulaPrime AI – Next-Gen Clinical Intelligence

What Inspired Us

Clinician burnout is one of the top challenges in today's healthcare. Each patient brings years of clinical records including lab results, imaging reports, admission/discharge records, surgical procedure, and treatment plan with corresponding results. Reviewing such large volume of multi-modal data quickly an making high-stake decisions adds significant cognitive load—especially under time pressure.

We imagined a system where AI agents could act like a team of junior specialists undertaking majority of the footwork and help our clinicians focus better on critical areas of treatment planning. Such Agentic systems can analyze large volume of complex, multi-modal data in parallel and detect and summarize key findings, so that clinicians wont have to look through everything. Rather, they can only validate the AI findings and probe deeper only in areas of interests.

This led to NebulaPrime AI—a prototype demonstrating how MultiAgentic AI, built using Google’s Agent Development Kit (ADK) to can deliver scalable, explainable clinical decision support system.

How We Built It

We implemented a three-tiered agent architecture using Google Cloud Platform and ADK:

1. Coordinator Agent

Determines the clinical context (e.g., oncology or nephrology)
Validates input data using Google Checks API for dangerous or explicit content as input guardrail
Routes the query to a relevant specialty coordinator (like oncology or nephrology)
Collects the Agentic output and checks the output for dangerous or profane content as output guardrail
Returns the data to requester

2. Specialty Coordinators

Oncology and Nephrology coordinators manage specialized disease assessment agents and runs them in parallel
Oncology agents: Breast cancer, colon cancer, brain cancer
Nephrology agents: Nephrolithiasis, renal failure, renal cell carcinoma
Puts their analysis output in context

3. Disease Assessment Agents

Each agent is focused in assessing a particular disease it is assigned
Each agent uses structured prompts and GenAI models (GPT-4o, Gemini Flash 2.0)
They analyze clinical reports, lab results, mammography results, biopsy results, and imaging reports
They use the provided thought process to assess whether patient's clinical data indicates a particular prognosis or lack of one
They use Google search tool to study further how to better detect the target disease or to understand the clinical data better
The formulate the prognosis with a confidence score
They also score whether the data was relevant or complete for the analysis
They return diagnosis, confidence scores, and rationale summaries in a structured format (JSON)

The backend system integrates API layers, data orchestration, and moderation, while the (not included) frontend allows clinicians to:

Upload patient context
Receive quick, interpretable disease assessments with insights

What We Learned

Agentic reasoning can mirror clinical workflows and specialist decision-making
Trust is built through confidence scoring and explainable summaries
Prompt engineering and data relevance filtering are core to reliable results
Guardrails like GCP Checks API are essential for ethical AI usage in healthcare
GenAI tools are powerful but need domain-specific constraints to be effective.

Challenges We Faced

Checks API is in private beta: While we found it useful, we did not enable it by default (code is in our github repo) in our public code base as it is likely to cause failures
Clinical data parsing: Free-text clinical reports required careful extraction and structuring
Prompt design: Ensuring each disease agent had accurate, scoped prompts was critical and time-consuming
MultiAgentic AI: Single layer of Agents loses accuracy, either loses details while working at strategical guidance or becomes too engaged in low level details an loses focus on strategic goal. MultiAgentic AI is critical in complex analysis. Performance guarantee with accuracy guarantee requires using all 3 forms of agent orchestration (free-form engagement, parallel, sequential)
Moderation sensitivity: Balancing safety filtering with valid medical terminology can be tricky as they contain both PII and PHI
Tool integration: Enabling live search for agents added depth but needed intelligent control and further guardrails as content from Internet can get injected now

Conclusion

NebulaPrime AI is not a production-ready clinical tool—it is a research prototype demonstrating how MultiAgentic AI can enable Next-Gen Clinical Intelligence. It shows the potential to dramatically reduce clinician prep time and cognitive load while maintaining transparency and trust.

We envision this as a foundational step toward real-world systems that ease cognitive burden, enhance diagnostic precision, and improve healthcare outcomes.