Enterprise Clinical Copilot

Enterprise Clinical Copilot (ECC) is a full-stack AI application. The application is designed to help healthcare professional assess unplanned readmission risk at discharge time for patients who have been admitted to the hospital for at least 24 hrs.

The application is built thoughtfully using trusted AI Architecture patterns. While the application itself is a prototype, the software is architected to run at enterprise scale, currently deployed on GCP.

This repo represents the backend system, where the agent, tools and automation to create those tools lives. The presentation layer is built using Django and is deployed through an app on my personal website. You can view the companion repo , to look at the source code for the web app.

The system is separated into deployable services and offline ML workflows:

enterprise_clinical_copilot/
├── services/
│   ├── agent/        LangGraph + Gemini HTTP service
│   └── mcp/          MCP server, prediction tool, retrieval tool, RAG pipeline
├── mlops/            Feature encoding, training, evaluation, and model serving
├── definitions/      Dataform ELT: cohort, splits, features, and marts
├── evaluation/       Agent, retrieval, and judge workflows
├── scripts/          Operational and data-generation commands
├── tests/            Offline service and pipeline tests
└── docs/             Architecture, contracts, evaluation, and operations

Runtime path

flowchart LR
	WEB["Django application"] --> AGENT["services/agent<br/>LangGraph + Gemini"]
	AGENT --> MCP["services/mcp<br/>MCP tools"]
	MCP --> MODEL["Vertex prediction endpoint"]
	MCP --> NOTES["BigQuery + Vector Search"]
	DEFINITIONS["Dataform definitions"] --> FEATURES["BigQuery features"]
	FEATURES --> MODEL

The Django application is maintained in the companion repository. It is the public boundary for authentication, demo-cohort authorization, quota, and presentation. The agent service owns language-model orchestration. The MCP service owns reusable prediction and retrieval tools.