Video summary

How to use AIGen: Automating AI Bill of Materials Generation Through Hybrid MLOps Integration

Main summary

Key takeaways

Technology

Overview

The subtitles describe AIGN (“AI gen”), a tool that automates generation of software/AI bill-of-materials (SBOM) metadata in the form of SPDX 3.x JSON-LD (referred to as “ABOM” and “SPDX 3A bombs”) using a hybrid MLOps integration approach centered around MLflow.

Core Capabilities and Workflow

1) Configuration-driven pipeline

  • Teams declare an artifact type (e.g., AI model, dataset).
  • Agents select the appropriate “builder” to extract metadata from the artifact’s source.

2) Credential handling

  • Sensitive information can be supplied via environment variables.

3) Metadata extraction

  • For MLflow models:
    • Retrieves model identity, version, timestamps, parameters, metrics, and SPDX-oriented (SBDX/SPDX-like) metadata
    • Source is an internal EMLflow/MLflow model registry
  • For datasets and other sources:
    • Uses corresponding builders (e.g., Kaggle, Hugging Face)

4) Serialization/output

After initializing builders and collecting metadata, AIGN:

  • Serializes the final ABOM as “SPDX 3 JSON-LD”

Generated output includes:

  • AI package entries for models (e.g., metrics, hyperparameters, versions, download locations)
  • Data set package entries (e.g., Iris dataset metadata, training context)

5) Relationship generation (SBDX)

The tool also creates SBDX/SPDX relationships, linking:

  • A model to its declared license
  • A model to the dataset it was trained on

Demonstrated Examples

1) MLflow + Kaggle

  • Source setup:
    • MLflow registry model (registered model in MLflow)
    • Kaggle dataset (example: Iris dataset)
  • Result: ABOM contains:
    • AI model package with MLflow-derived metadata
    • Dataset package with Kaggle dataset metadata
    • Relationships:
      • model ↔ license
      • model ↔ training dataset

2) Reused pipeline with Hugging Face instead

  • Change:
    • Instead of querying a local MLflow registry, the tool uses a Hugging Face builder to retrieve metadata from a public model hub.
  • Extracted from Hugging Face model cards (when available):
    • Model identity, download location, supplier
    • Release timestamps, model type, domain
    • License information
    • Textual metadata from the model card
  • Result:
    • The same overall ABOM generation pipeline is reused
    • Only builder configuration changes
    • Output remains SPDX 3A / “SBDX 3A” style with machine-readable metadata

Key Benefits Emphasized

  • Generates SPDX 3.x-style machine-readable metadata from heterogeneous artifact sources, including:
    • MLflow registries
    • Kaggle datasets
    • Public model hubs (Hugging Face)
  • Intended to support:
    • ML supply chain transparency
    • documentation
    • downstream compliance workflows

Main Speakers / Sources (as stated in subtitles)

  • AIGN / “AI gen” (demo subject)
  • MLflow (internal model registry source)
  • Kaggle (dataset source)
  • Hugging Face (public model hub source)
  • SPDX / SBDX (metadata/relationship formats referenced)

Original video