/ml-build

Read the handoff document (if provided) or ask the user to describe the problem.

1. Read agents/model-architect.md for architecture design guidance.
2. Read the relevant reference files based on the architecture type:
   • Transformers/LLMs: references/transformers-patterns.md
   • GNNs/graphs: references/gnn-cookbook.md
   • General PyTorch: references/pytorch-patterns.md
   • RL: references/advanced-systems.md
3. Check templates/ for a matching starting point:
   • Standard training: templates/train-loop.py
   • GNN pipeline: templates/gnn-pipeline.py
   • LoRA fine-tuning: templates/fine-tune-lora.py
   • KGE: templates/kge-pipeline.py
   • RL: templates/rl-agent.py
4. Produce these files:
   • model.py - nn.Module with config dataclass, shape comments, smoke test
   • dataset.py - Data loading with verification mode
   • train.py - Complete training pipeline (AMP, grad clip, checkpointing)
   • config.py - All hyperparameters in dataclasses
5. Run the quality gate checklist from CLAUDE.md before presenting code.