mloda-demo

Live demo built for the applydata Berlin 2026 talk "Building Deterministic Context Layers for AI Agents".

Mixed-source credit-risk pipeline: JSON + synthetic Excel + synthetic Markdown → one row per customer → an MLP classifier trained on UCI German Credit → Fraunhofer Zennit LRP attribution → method swap (EpsilonPlus ↔ Gradient). All orchestrated by mloda FeatureGroups. Executable as a deterministic CLI tool.

The point of the demo: the LLM / agent on top is non-deterministic. The context layer below it is deterministic. Same mixed-source inputs always produce the same predictions and the same explanations.

Quick start

uv venv
source .venv/bin/activate
uv sync --all-extras

which mloda-demo  # verify CLI is installed
mloda-demo discover
mloda-demo run --customer app-customer-c
mloda-demo predict --customer app-customer-c
mloda-demo explain --customer app-customer-c

See demo/applydata_handbook.md for the full 6-act demo script.

Run all checks with tox. Run integration tests with pytest -m slow.

Structure

mloda_demo/
├── feature_groups/
│   ├── inputs/                   # 3 root FGs: applications.json, xlsx, markdown
│   └── classifier/               # MLP + artifact + CreditRiskClassifierFG
├── xai/
│   ├── attribution/              # Zennit LRP + Gradient attribution FGs
│   └── visualization/            # heatmap renderer
demo_data/                        # customer data + trained artifacts
demo/                             # applydata_handbook.md (CLI demo script)
tests/                            # unit + integration CLI tests

Open Source Libraries

• mloda — feature orchestration framework (Apache 2.0)
• mloda-registry — plugins, guides, and best practices (Apache 2.0)
• PyTorch — MLP model training and inference (BSD 3-Clause)
• Zennit — Layer-wise Relevance Propagation (LGPLv3+)
• OpenML — German Credit dataset source
• mloda-plugin-template — starting point for this repo