version: "1.0.0" name: ciel-ml-and-data-patterns description: CIEL's framework for PyTorch ML patterns and Postgres data optimization. license: MIT metadata: ciel-version: 1.0.0 ciel-extension: ciel.yaml

CIEL ADAPTATION: ML & Data (The Intelligence Layer)

This skill manages high-performance data patterns, from deep learning training loops to SQL query optimization.

PyTorch ML Patterns

1. Device-Agnostic: ALWAYS use device = torch.device(...). Prohibit hardcoded .cuda() calls.
2. Reproducibility: Set seeds for torch, np, and random in a central set_seed() function.
3. Shape Integrity: Annotate and verify tensor shapes in the forward() pass comments.
4. Efficiency: Use optimizer.zero_grad(set_to_none=True) and model.eval() for validation.

Postgres Data patterns

• Indexing: Equality columns first, then range columns. Use GIN for JSONB and BRIN for time-series.
• Types: Use bigint for IDs, timestamptz for times, and text for variable strings.
• Pagination: Use keyset/cursor pagination (WHERE id > $last_id) instead of OFFSET for O(1) performance.
• Security: Wrap RLS policies in (SELECT auth.uid()) = user_id for optimization.

Memory Management

• AMP: Use torch.amp.GradScaler for mixed-precision performance.
• Checkpointing: Save model_state_dict AND optimizer_state_dict to allow resuming training.

Anti-Patterns

• In-place Mutation: Using x += residual in PyTorch (breaks autograd). Use x = x + residual.
• Small Inserts: Performing SQL inserts in a loop instead of batching.
• Select \*: Reading every column in SQL (causes unnecessary I/O bloat).