Real-Time FDA Clinical Data Monitor using AutoGen and Gemini
Inspired by the FDA pilot program for real-time clinical data feeds, this challenge tasks you with building a multi-agent system to monitor and analyze pharmaceutical trial data. You will use AutoGen to orchestrate a team of agents that process high-frequency clinical updates, identify safety signals, and report anomalies. The system must utilize Gemini 3.1 Flash Lite for rapid data extraction and Sweep AI to automatically manage code patches for the data ingestion pipelines. To ensure model reliability, you will integrate deepchecks for continuous evaluation of the agent outputs. Finally, you will design a low-code dashboard in Bubble to visualize these real-time feeds and use Cartesia for voice-enabled AI alerts when critical safety thresholds are met.
What you are building
The core problem, expected build, and operating context for this challenge.
Build an autonomous multi-agent clinical monitoring system using AutoGen and Gemini 3.1 Flash Lite for real-time FDA data analysis.
How work is evaluated
Evaluation focuses on the system's ability to correctly identify safety signals in simulated real-time data and the performance of the Sweep AI auto-patching mechanism.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
Latency Test
System must process a data batch in under 2 seconds
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Signal Accuracy
Percentage of correctly identified clinical anomalies • target: 0.95 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Master AutoGen for building stateful multi-agent conversations with specific roles for data scientists and safety monitors
Orchestrate Gemini 3.1 Flash Lite for low-latency extraction of clinical entities from unstructured real-time feeds
Integrate Sweep AI to autonomously handle GitHub issues and pull requests for clinical data parsers
Implement Cartesia voice synthesis for low-latency critical alerts in the monitoring interface
Build a real-time data visualization bridge between Python-based AutoGen and Bubble via API Connector
Deploy deepchecks to perform validation of LLM outputs against ground-truth clinical safety protocols
Reference links and supporting material
Sample data for 1 tasks
Challenge setup
Evaluation focuses on the system's ability to correctly identify safety signals in simulated real-time data and the performance of the Sweep AI auto-patching mechanism.
Challenge input
JSON stream of clinical observation records
Google Gemini
Google's multimodal AI model
AutoGen
Agent Systems · Multi-Agent Systems
deepchecks
ML testing and monitoring
Evaluated output
JSON Alert with severity level and reasoning
- System must process a data batch in under 2 seconds
- Percentage of correctly identified clinical anomalies • target: 0.95 • range: 0-1
- Signal Accuracy target: 0.95
- 1 public reference case
- Python execution harness
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- Google Gemini · Required
- AutoGen · Optional
- deepchecks · Optional
Evaluation contract
- Latency Test · Weight 1
- Signal Accuracy · Weight 1
Metrics
- Signal Accuracy · Target 0.95
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON stream of clinical observation records
Output: JSON Alert with severity level and reasoning
Execution harness: Python
Recipe freeze marker
No recipe freeze time is recorded.