The conversation

  1. 01An agent's trust boundary: which entities it may read, and which actions it may take
  2. 02An agentic step a product can commit, scoped to one entity, with an audit and a rollback
  3. 03RAG as a contract: governed sources, a freshness bound, and a citation back to the record
  4. 04LLM context cut to the decision, and a hard stop when the evidence runs out
  5. 05Voice held to the same entity a person can open and check
  6. 06Resolve the entity and retrieve the evidence before the model is allowed to generate
  7. 07Tool calling inside a blast radius the platform can audit and revoke
  8. 08A model may propose a score. The platform is what commits it.
  9. 09Risk on a business profile, assessed as two layers: the network, and the application
  10. 10That assessment as a platform, built in four to five days with Claude on the implementation and the scoring contract held in the architecture
  11. 11Company profiles at scale, every risk signal traced to its source
  12. 12Redshift held as separate contracts: landing, the model, and the mart
  13. 13One metadata definition that runs Spark, Python, and dbt
  14. 14Kinesis and Firehose into S3, live traffic and replay on the same field-level rules
  15. 15Redshift on the curated prefix as the query surface, Lake Formation where a subject must be deleted
  16. 16A regional copy that already carries its residency, its PII class, and its latency budget
  17. 17PII under data-protection law, enforced on the path: purpose, access, and deletion of every copy
  18. 18Cyber risk as evidence a decision can defend
  19. 19One entity question across firmware, chips, mobility, supply chain, banking, and healthcare
  20. 20A master record that products, marts, and agents all resolve to