relflow

relflow builds PyTorch models from nested records. Fields describe values; branches describe relationships. A customer can contain orders, and each order can contain line items. Those local contexts feed a shared representation from which the model predicts missing or supervised values.

tenure_days: 420
orders:
  - age_days: 3
    line_items:
      - sku: A12
        quantity: 2
      - sku: B07
        quantity: 1
churned: false
next_month_spend: 75.0

Applications supply Arrow tables or Polars frames.

Define your first model →

Learn The Shape

• Model tree: fields, local contexts, and predictions.
• Data types: choose a representation for each value.
• Online vocabulary: learn labels and handle new or unknown values.
• Binding data: connect record keys to schema names.
• Masking: decide what the model sees and learns.

Use The Model

• Preprocessing: prepare fields and nested records with Polars.
• Training: fit, validate, save, and restore with Lightning.
• Pretraining and fine-tuning: transfer learned context to a labeled task.
• Prediction output: read typed results and embeddings.
• Customer risk: combine nested histories and targets.

Explore The Evidence

Proofs show the schemas, controls, and recorded results behind individual learning claims, including known limitations. Start with a task close to yours and inspect what its model can observe.

Use this hierarchy when the relationships carry useful information. When a flat feature table already expresses the task, a simpler model may suffice.