Patterns from the work

The model proposes. Deterministic controls decide what it is allowed to change. A human validation gate releases the action, and the audit trail keeps the query, the sources, the proposal, and the person who let it through. Taken to production.

What it changed: A TPRA manual review that took about 30 hours now takes about 10 minutes of end-to-end processing time, and every finding still passes a human reviewer before release.

Customer-facing agents for outage, billing, and service — with guardrails against hallucinated grid information and full regulatory audit trails.

What it changed: Outage, billing, and service answers stay inside a boundary, and the audit trail is what a review uses when an answer is disputed.

A personal reference build for source-verified policy retrieval — every answer cited, every query logged, drift monitored. Not a client engagement.

What it changed: An answer is withheld unless it cites a source, and every query is logged so a challenge can be checked against the record.

A personal reference build for the operational layer production agents require — LLM tracing, evaluation pipelines, drift detection, and cost. Not a client engagement.

What it changed: A single run leaves a trace, an evaluation, and a cost, so a bad answer can be found without reconstructing the call.

Observability and anomaly detection in a large enterprise environment: predictive incident management, with operations and engineering in the same conversation. I am not publishing figures from that work.

What it changed: Incidents that used to surface late became visible from telemetry, while production uptime stayed the constraint on what could change.