SilentSurge

System architecture

Inspiration

Sepsis kills 11 million people per year, and 80% of those deaths are preventable with early detection. ICU patients cannot advocate for themselves when their condition deteriorates silently. We built SilentSurge because every minute without detection is a minute closer to organ failure.

What it does

SilentSurge is an autonomous multi-agent ICU monitoring system trained on real MIMIC-III clinical data. Four specialized agents run in parallel:

VitalsWatcher — monitors heart rate, temperature, blood pressure, and lactate trends

LabInterpreter — analyzes lab panels for sepsis biomarkers

MedReviewer — detects beta-blocker masking, where medications artificially suppress heart rate and hide true sepsis severity

EscalationAgent — triggers a physician briefing via LLaMA 3.3 70B on Groq when a patient reaches critical risk

A LoopAgent performs iterative self-correction to catch masked cases. High-risk patients require human-in-the-loop physician approval before escalation — AI assists, physicians decide.

How we built it

Google ADK — ParallelAgent runs all specialist agents simultaneously; LoopAgent handles self-correction

Gemini 2.5 Flash — powers agent reasoning and dynamic risk scoring

Groq + LLaMA 3.3 70B — generates real-time physician briefings in seconds

MIMIC-III — real de-identified ICU data across ChartEvents, LabEvents, and Prescriptions

Streamlit — human-in-the-loop approval dashboard

Google Cloud Run — live cloud deployment

Challenges we ran into

Beta-blocker masking was the hardest clinical problem — patients on these medications show artificially normal heart rates, which standard sepsis scoring misses entirely. We built a dedicated self-correction loop to detect and adjust for this.

Managing parallel API calls across 93 patients without hitting Groq or Gemini rate limits

Deploying a Streamlit app with WebSocket support on Cloud Run required careful configuration

Accomplishments that we're proud of

Successfully integrating Google ADK's ParallelAgent and LoopAgent in a real medical context

Detecting the subtle clinical edge case of beta-blocker masking using agentic self-correction

Generating coherent, physician-ready clinical briefings autonomously via LLaMA 3.3 70B

End-to-end deployment from raw MIMIC-III data to a live Cloud Run application

What we learned

Google ADK's multi-agent architecture is genuinely powerful for real-time triage. More importantly, we learned that AI in healthcare must augment, never replace, physician judgment — which is why human-in-the-loop approval is a core feature, not an afterthought.