Writing & Thinking

Notes on compliant AI

I’m distilling what I’ve learned from building, and defending, AI systems inside a regulated firm into working notes for compliance and engineering teams: how to architect a compliant AI system, from rule-set design to evals to tracing.

Topics in progress:

  • Rule sets that police verdicts, not vocabulary
  • Evals as a supervisory control (and when a false-negative rate is a finding)
  • The audit trail as the unit of accountability
  • Why “deterministic pipeline, probabilistic component” beats “aligned model”
  • Supervision by design: staffing a two-person team like a twenty-person one

Want to be notified when these publish? Drop me a note.