Use Cases
Data workflows
- Training and evaluation data: extract metadata, curate a selection and feed PyTorch.
- Physical AI and robotics: curate LeRobot camera episodes or select MCAP recordings by sensor topics.
- Data exploration: query saved metadata and model outputs to investigate failures, find rare conditions and compare data coverage. See data operations.
- Large inventories: discover objects across S3, GCS and Azure from storage metadata. See multi-cloud inventory.
Agent workflows
Patterns for building on saved results across sessions, with examples and code.
- Agent compounding across sessions: each session reads typed datasets the last session produced, instead of re-deriving from raw bytes.
- Agent anti-patterns over object storage: the things coding agents do wrong on S3 by default, and the chain pattern that replaces each one.
- Agent as a dataset producer: the agent's output is a versioned typed dataset, not a notebook artifact. Other agents and people query it by name.
- Retroactive agent runs: a new model drops; re-run it across yesterday's frames without a full recompute, paying only for the new work.
- Cross-agent reasoning: Claude Code's saved dataset becomes Cursor's input. The Knowledge Base is the contract.