NVIDIA · NCP-AAI

The oversight thread

Where a human actually stands in the loop. Opens in M1 with the human-agent interface as the architectural oversight surface, runs through M9's guardrail rail stages and PII/compliance controls, and closes in M10's human-in-the-loop techniques, structured feedback loops, transparency, and named intervention points.

NCPA-T4 · 18 lessons across 3 modules

  1. M1M1-01Agent architecture styles: reactive, deliberative, and hybrid systems
  2. M1M1-02Structuring multi-step reasoning: logic trees, prompt chains, and adaptable architecture
  3. M1M1-03ReAct: interleaving reasoning and acting
  4. M1M1-04Multi-agent orchestration topologies: centralized, decentralized, federated, and hierarchical
  5. M1M1-05Memory as an architectural concern: short-term vs. long-term
  6. M1M1-06Knowledge graphs for relational, multi-hop reasoning
  7. M1M1-07Designing the human-agent interface as the oversight surface
  8. M9M9-01NeMo Guardrails: the five rail stages
  9. M9M9-02Layered safety frameworks: filters and escalation
  10. M9M9-03PII, agentic security, and audit trails
  11. M9M9-04Mitigating bias and toxicity in agent outputs
  12. M9M9-05Licensing and regulatory compliance
  13. M10M10-01Human-in-the-loop: benefits and drawbacks
  14. M10M10-02HITL techniques: supervised learning, RLHF, and active learning
  15. M10M10-03Structured feedback loops for iterative improvement
  16. M10M10-04Transparency and decision traceability
  17. M10M10-05Human oversight and intervention points
  18. M10M10-06Intuitive UI as the oversight surface

Part of the throughlines running across the NCP-AAI prep course.