Olympus

Clinical Trial Data Pipeline with AI-powered Code Generation.

Overview

Olympus implements a Medallion architecture for clinical trial data with unified PydanticAI agents:

┌─────────┐    ┌─────────┐    ┌─────────┐    ┌───────────┐
│  Meta   │───▶│  Bronze │───▶│  Silver │───▶│   Gold    │
│ (Spec)  │    │  (Raw)  │    │(Derived)│    │ (Stats)   │
└─────────┘    └─────────┘    └─────────┘    └───────────┘
     │              │              │              │
     └──────────────┴──────────────┴──────────────┘
                          │
                   ┌──────▼──────┐
                   │  Platinum   │
                   │ (Slide Deck)│
                   └─────────────┘

Installation

# Clone and setup
git clone <repo-url>
cd olympus
uv sync

# Set up environment variables
cp .env.example .env
# Edit .env and add your API keys (DEEPSEEK_API_KEY, ZHIPU_API_KEY)

Quick Start

Olympus provides an MCP Server for use with OpenCode, Claude Desktop, and other MCP-compatible AI tools.

Setup:

# Clone and setup
git clone <repo-url>
cd olympus
uv sync

# Configure OpenCode
cat > ~/.config/opencode/config.json << 'EOF'
{
  "mcpServers": {
    "olympus": {
      "command": "uv",
      "args": ["--directory", "/path/to/olympus", "run", "olympus-mcp"]
    }
  }
}
EOF

# Restart OpenCode
opencode

Usage Examples:

User: 帮我列出some_study的deliverables

OpenCode: [调用 list_mock_deliverables tool]
         Found 34 deliverables:
         - 14.1.1 CAR-T Treatment Summary
         - 14.1.2.1 Baseline Characteristics (IIM Cohort)
         ...

User: 提取mock shell到JSON

OpenCode: [调用 extract_mock_shell tool]
         Extracted to mock_extracted.json
         Deliverables: 34

User: 加载metadata到数据库

OpenCode: [调用 load_meta tool]
         Loaded to study.duckdb:
           Silver variables: 45
           Params: 12

User: 生成Silver transforms

OpenCode: [调用 generate_silver tool]
         Generated 23 transforms for baseline:
           High confidence: 18
           Medium confidence: 5

User: 生成PPTX

OpenCode: [调用 generate_platinum tool]
         Generated slide deck:
           Slides: 67
           Deliverables: 34

Available MCP Tools (15):

Category	Tools
Meta	load_meta, extract_mock_shell, list_mock_deliverables, list_db_deliverables, get_variable_details
Silver	generate_silver, list_silver_transforms, get_transform_code, get_meta_variables
Gold	generate_gold, list_gold_transforms
Platinum	generate_platinum, preview_platinum_deliverable
Utility	lookup_als_field, get_bronze_schema

See AGENTS.md for detailed tool documentation.

Project Structure

olympus/
├── src/olympus/
│   ├── core/           # Database, schema models, Pydantic models
│   ├── agent/          # LLM providers (zhipu, deepseek) + PydanticAI base
│   ├── meta/           # Metadata generation (agent, generator, loader)
│   ├── bronze/         # Raw data import (SAS/CSV/Excel)
│   ├── silver/         # Silver agent + transformation engine
│   ├── gold/           # Gold agent + statistics engine
│   ├── platinum/       # Platinum agent + PPTX renderer
│   └── transforms/     # Python transformations registry
├── tests/
├── examples/
└── archive/            # Deprecated code

Data Layers

Layer	Schema	Description
Meta	meta.*	9 metadata tables
Bronze	bronze.*	Raw EDC data (one table per form)
Silver	silver.baseline/longitudinal/occurrence	Subject-level derived data
Gold	gold.baseline/longitudinal/occurrence	Group-level statistics
Platinum	-	PowerPoint slide decks

Meta Schema (9 Tables)

Table	Purpose
meta.study_info	Study information
meta.visits	Analysis visit definitions
meta.form_classification	Form → Domain/Schema mapping
meta.params	Longitudinal parameter definitions
meta.attrs	Occurrence domain extended fields
meta.bronze_dictionary	Bronze layer data dictionary
meta.silver_dictionary	Silver layer data dictionary
meta.gold_dictionary	Gold layer statistics definitions
meta.platinum_dictionary	Platinum deliverable definitions

PydanticAI Architecture

All layers use unified PydanticAI agents:

┌─────────────────────────────────────────────────────────┐
│                    PydanticAI Agent                      │
│                                                          │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐ │
│  │   Meta   │  │  Silver  │  │   Gold   │  │ Platinum │ │
│  │  Agent   │  │  Agent   │  │  Agent   │  │  Agent   │ │
│  └────┬─────┘  └────┬─────┘  └────┬─────┘  └────┬─────┘ │
│       │             │             │             │       │
│       └─────────────┴─────────────┴─────────────┘       │
│                          │                               │
│                   ┌──────▼──────┐                       │
│                   │ Shared Tools│                       │
│                   │ - ALS lookup│                       │
│                   │ - Meta vars │                       │
│                   │ - Deps check│                       │
│                   └─────────────┘                       │
└─────────────────────────────────────────────────────────┘

LLM Providers

Provider	Usage
DeepSeek (default)	--provider deepseek
Zhipu	--provider zhipu

Set your API key:

export DEEPSEEK_API_KEY=your_key_here
# or
export ZHIPU_API_KEY=your_key_here

Development

# Run tests
uv run pytest tests/ -v

# Format code
uv run --extra dev black src/olympus/

# Lint code
uv run --extra dev ruff check src/olympus/

# Type check
uv run --extra dev mypy src/olympus/

Documentation

• AGENTS.md - Coding agent guidelines and MCP tools reference
• ARCHITECTURE.md - Detailed architecture design

License

MIT