Clinical Supply Chain Agentic Data Platform

AI-powered platform that validates, explains, and remediates clinical supply data quality across Bronze → Silver → Gold on Azure Databricks, with governed MCP tools, data contracts, and an evaluation harness mapped to G7 Engineer II criteria T1 / T2 / T3.

Storage: Databricks only — Unity Catalog Volumes for landing files and managed Delta tables for Bronze/Silver/Gold. No separate Azure Data Lake / ADLS account in this design.

Quick start (local)

python -m venv .venv
# Windows
.venv\Scripts\activate
pip install -r requirements.txt

# List contracts
python -m clinical_platform.cli list-contracts

# Happy-path validation
python -m clinical_platform.cli validate sap_batch_raw

# Portfolio demo — schema drift
python -m clinical_platform.cli demo-drift

# Agents
python -m agents.cli dq --drift
python -m agents.cli rca
python -m agents.cli ask "Why did Gold trial inventory fail today?"

# T3 evaluation
python -m eval.harness

Copy .env.example → .env when you wire Azure OpenAI (offline fallback works without it).

Repository layout

contracts/           # YAML table contracts (source of truth)
clinical_platform/   # validate_table, expectations, local engine
mcp_server/          # Governed tools for agents
agents/              # DQ, RCA, Q&A
prompts/             # Versioned system prompts
eval/                # Golden Q&A + harness
metadata/ddl.sql     # Unity Catalog metadata tables
notebooks/           # Databricks job entrypoints
workflows/           # Databricks Asset Bundle stub
docs/                # Architecture notes
tests/               # Unit tests

Mapping to assessment criteria

Criterion	What this repo proves
T1	MCP tool server, reusable agent skills, prompt-backed RCA/Q&A, knowledge via contracts/lineage
T2	Contract-driven validation, quality reports, lineage stubs, GxP-gated remediation proposals
T3	Golden eval set, keyword/tool groundedness scores, versioned prompts, nightly job stub

Databricks next steps

1. Run metadata/ddl.sql in your workspace (creates catalogs, schemas, Volumes, metadata tables).
2. Sync contracts into clinical_supply.metadata.table_contracts.
3. Land source files in /Volumes/clinical_supply/bronze/landing/ (not ADLS).
4. Point notebooks at cluster libraries (pip install -e .).
5. Deploy workflows/databricks.yml with Databricks Asset Bundles.
6. Replace local engine with Spark/Great Expectations runner when ready.