Dhruv Nathawani introduces NVIDIA Data Designer as a way to turn desired agent behavior into a reproducible data pipeline. Seed data and deterministic samplers supply topics, difficulty and other variation; model-generated columns then produce questions, answers or structured outputs. Small previews support inspection before larger batches.
Dhruv Nathawani builds from question-answer pairs to search-tool trajectories, using MCP to let a model retrieve evidence rather than answer only from memory. For harder search riddles, sampled paths through Wikidata connect several entities, encouraging multi-step reasoning and tool use. The resulting examples can support supervised training, while prompt-only datasets can feed a separate reinforcement-learning workflow.
Dhruv Nathawani emphasizes filtering and layered evaluation rather than assuming synthetic answers are correct. Stale seed facts, weak matching rules and imperfect model judges all create risks; human review and current evidence provide additional checks. He also discusses demographic personas, structured-output tasks and the importance of diverse, task-appropriate data over raw volume.
Watch on YouTube
