CardioMAS — Cardio Multi-Agent System

A locally-runnable multi-agent system that analyzes ECG datasets and generates reproducible train/validation/test splits. Outputs are saved locally by default. Publishing to vlbthambawita/ECGBench on HuggingFace is an explicit opt-in step that requires write access.

Architecture

flowchart TD
    CLI(["`**CLI / Python API**
    cardiomas analyze`"])
    CLI --> ORCH

    subgraph PIPELINE ["LangGraph Pipeline — V2 Hub-and-Spoke"]
        ORCH(["🎯 **Orchestrator**
        Dynamic routing hub
        sets next_agent each turn"])

        %% Worker agents (spokes)
        NL["📝 NL Requirement
        Parse natural-language
        user requirements"]

        DISC["🔍 Discovery
        Identify dataset type,
        source & metadata"]

        PAPER["📄 Paper
        Find & parse paper
        Extract split methodology"]

        ANALYSIS["📊 Analysis
        Scan files, parse CSV
        Compute statistics"]

        SPLIT["✂️ Splitter
        SHA-256 seeded
        deterministic splits"]

        SEC["🔒 Security
        PII scan · raw-data
        Patient leakage check"]

        CODER["💻 Coder
        Generate custom
        split code"]

        PUB["☁️ Publisher
        Push to HF
        Update GitHub README"]

        %% Orchestrator → workers (dispatch)
        ORCH -->|dispatch| NL
        ORCH -->|dispatch| DISC
        ORCH -->|dispatch| PAPER
        ORCH -->|dispatch| ANALYSIS
        ORCH -->|dispatch| SPLIT
        ORCH -->|dispatch| SEC
        ORCH -->|dispatch| CODER
        ORCH -->|dispatch| PUB

        %% Workers → Orchestrator (return)
        NL -->|done| ORCH
        DISC -->|done| ORCH
        PAPER -->|done| ORCH
        ANALYSIS -->|done| ORCH
        SPLIT -->|done| ORCH
        SEC -->|done| ORCH
        CODER -->|done| ORCH
        PUB -->|done| ORCH

        %% Terminal routes from orchestrator
        ORCH -->|cache hit| EXISTING["↩ Return existing\nsplits from HF"]
        ORCH -->|audit failed| ERR["❌ End with error\nBlocked — not saved"]
        ORCH -->|pipeline done| SAVED["💾 End — saved locally"]
    end

    PUB -.->|if --push| HF[("HuggingFace
    vlbthambawita/ECGBench")]
    SAVED --> OUT

    OUT["`**output/<dataset>/**
    splits.json
    split_metadata.json
    analysis_report.md`"]

    style PIPELINE fill:#1a1a2e,stroke:#4a9eff,color:#fff
    style ORCH fill:#0f3460,stroke:#4a9eff,color:#fff,shape:stadium
    style NL fill:#2d1b4e,stroke:#9b59b6,color:#fff
    style DISC fill:#1b4332,stroke:#40916c,color:#fff
    style PAPER fill:#3b1f5e,stroke:#9b59b6,color:#fff
    style ANALYSIS fill:#7b4220,stroke:#e67e22,color:#fff
    style SPLIT fill:#1a5276,stroke:#2e86c1,color:#fff
    style SEC fill:#641e16,stroke:#e74c3c,color:#fff
    style CODER fill:#1a3a4a,stroke:#00bcd4,color:#fff
    style PUB fill:#145a32,stroke:#27ae60,color:#fff
    style SAVED fill:#1a3a1a,stroke:#27ae60,color:#fff
    style ERR fill:#3b0f0f,stroke:#e74c3c,color:#fff
    style HF fill:#2d2d2d,stroke:#f5a623,color:#fff
    style OUT fill:#2d2d2d,stroke:#4a9eff,color:#fff
    style EXISTING fill:#2d2d2d,stroke:#4a9eff,color:#fff
    style CLI fill:#0d0d0d,stroke:#4a9eff,color:#fff

Each node is a dedicated LLM-backed agent. The orchestrator is the central hub — it dynamically decides which agent to invoke next after each agent completes (hub-and-spoke pattern). Every worker agent returns to the orchestrator after finishing. Agents communicate only through the shared GraphState. The security agent is a hard gate — publishing is blocked if any check fails.

Requirements

• Python ≥ 3.10
• Ollama running locally with a model pulled (default: gemma4:e2b)

ollama pull gemma4:e2b
ollama serve
pip install cardiomas

Quick Start

# Analyze a dataset — results saved to ./output/ptb-xl/
cardiomas analyze https://physionet.org/content/ptb-xl/1.0.3/

# Use a local directory
cardiomas analyze /data/ptb-xl/

# Stream all agent reasoning live
cardiomas analyze /data/ptb-xl/ --verbose

# Analyze and push to HuggingFace in one step (requires HF_TOKEN)
cardiomas analyze /data/ptb-xl/ --push

After analyze, the following files are written locally:

output/
└── ptb-xl/
    ├── splits.json           # train/val/test record IDs + reproducibility config
    ├── split_metadata.json   # seed, strategy, version, timestamp
    └── analysis_report.md    # LLM-generated dataset analysis

CLI Reference

cardiomas analyze

Analyze a dataset and save splits locally.

cardiomas analyze DATASET_SOURCE [OPTIONS]

Verbose output

--verbose (-v) prints every agent step and LLM prompt/response in real time instead of showing a spinner:

cardiomas analyze /data/ptb-xl/ -v

Verbose mode on — streaming agent output below.

  [orchestrator] pipeline start — source: /data/ptb-xl/
  [orchestrator] no cache hit — running full pipeline
  [discovery]    registry hit → ptb-xl
  [paper]        searching arXiv: 'ptb-xl ECG dataset electrocardiogram'
  [paper]        found 3 result(s)
  [paper]        calling LLM (ChatOllama)…
──────────────── paper — LLM call ────────────────
╭─ prompt ────────────────────────────────────────╮
│ Analyze this ECG dataset paper and extract: …  │
╰─────────────────────────────────────────────────╯
╭─ response ──────────────────────────────────────╮
│ 1. Official splits: Yes (Section 2.3, page 4)  │
╰─────────────────────────────────────────────────╯
  [analysis]     found 42 files
  [splitter]     saved splits → output/ptb-xl/splits.json
  [security]     audit PASSED — no PII, no raw data, no leakage

Each agent is color-coded. Without --verbose, only a spinner runs during the pipeline and a summary table is shown at the end.

cardiomas push

Push previously saved local splits to HuggingFace. Requires HF_TOKEN.

cardiomas push ptb-xl
cardiomas push ptb-xl --output-dir /my/results

Runs a security audit before uploading. Refuses to push if any check fails.

cardiomas status

Check if a dataset has published splits on HuggingFace.

cardiomas status ptb-xl

cardiomas list

cardiomas list              # show known datasets (registry)
cardiomas list --remote     # show datasets published on HuggingFace
cardiomas list --local      # show locally cached datasets

cardiomas verify

Re-check reproducibility metadata of published splits.

cardiomas verify ptb-xl --seed 42

cardiomas contribute

Submit community splits to vlbthambawita/ECGBench.

cardiomas contribute ptb-xl --split-file my_splits.json

cardiomas config

cardiomas config --show
cardiomas config --set OLLAMA_MODEL=mistral

HuggingFace Publishing (opt-in)

Publishing requires write access to vlbthambawita/ECGBench. Set your token before pushing:

export HF_TOKEN=hf_...
cardiomas push ptb-xl

Only record identifiers are ever published — no raw ECG signals, no patient data.

Python API

from cardiomas import CardioMAS

mas = CardioMAS(ollama_model="gemma4:e2b", seed=42)

# Analyze and save locally
result = mas.analyze("/data/ptb-xl/")
print(result["local_output_dir"])   # output/ptb-xl

# Analyze and push to HuggingFace
mas.analyze("/data/ptb-xl/", push_to_hf=True)

# Read back published splits
splits = mas.get_splits("ptb-xl")
train_ids = splits["train"]

# Custom splits
mas.analyze(
    "/data/ptb-xl/",
    custom_split={"train": 0.7, "val": 0.15, "test": 0.15},
    stratify_by="scp_codes",
    seed=123,
)

Using Different Local Models

Any model available in Ollama works. Pull a model, then point CardioMAS at it:

# Default (recommended for full pipeline)
ollama pull gemma4:e2b

# Larger Gemma 4 variants for heavier reasoning tasks
ollama pull gemma4:e4b
ollama pull gemma4:27b

# DeepSeek Coder — best for the coding agent
ollama pull deepseek-coder:6.7b

# Use a specific model for the whole pipeline
OLLAMA_MODEL=gemma4:e2b cardiomas analyze /data/ptb-xl/

# Or set it permanently in .env
echo "OLLAMA_MODEL=gemma4:e2b" >> .env

Per-Agent LLM Configuration

Available in v0.2.0 (dev/v2-dynamic-orchestrator)

Each agent can use a different LLM. This is useful when you want a fast, lightweight model for simple tasks (discovery, security scan) and a more capable model for reasoning-heavy tasks (analysis, coding).

Via environment variables

# Fallback for all agents
OLLAMA_MODEL=gemma4:e2b

# Per-agent overrides (all optional)
AGENT_LLM_ORCHESTRATOR=gemma4:e2b
AGENT_LLM_NL_REQUIREMENT=gemma4:e2b
AGENT_LLM_DISCOVERY=gemma4:e2b
AGENT_LLM_PAPER=gemma4:e2b
AGENT_LLM_ANALYSIS=gemma4:e2b
AGENT_LLM_SPLITTER=gemma4:e2b
AGENT_LLM_SECURITY=gemma4:e2b
AGENT_LLM_CODER=deepseek-coder:6.7b
AGENT_LLM_PUBLISHER=gemma4:e2b

Set these in .env or export them before running cardiomas.

Via CLI flags

cardiomas analyze /data/ptb-xl/ \
  --llm-coder deepseek-coder:6.7b \
  --llm-analysis gemma4:e4b \
  --llm-discovery gemma4:e2b

Via Python API

from cardiomas import CardioMAS

mas = CardioMAS(
    agent_llms={
        "coder":    "deepseek-coder:6.7b",
        "analysis": "gemma4:e4b",
        "default":  "gemma4:e2b",   # fallback for all other agents
    }
)
mas.analyze("/data/ptb-xl/")

Model recommendations

Verbose LLM name display

With --verbose, each LLM call shows the model name and backend:

──────────── paper — LLM call [gemma4:e2b @ ollama] ────────────

Environment Variables

Copy .env.example to .env and fill in as needed.

Links

• HuggingFace Dataset: vlbthambawita/ECGBench
• PyPI: cardiomas
• GitHub: vlbthambawita/CardioMAS