Brevor Research.
Findings on behavioral-AI failure modes, audit primitives, and cultural context. Open citations. Closed product code. Funded by Brevor Tech with 1% of operating revenue.
Behavioral routing at sub-100ms: lessons from 1.8B production calls.
Architecture notes from running the routing engine at production load. What broke, what scaled, what we'd build differently.
Latency budgets for behavioral routing.
The behavioral layer adds round-trip cost. How to budget it: target latencies by use case, fast-paths for low-risk prompts, hot-swappable context.
Cultural modules: why one global model fails healthcare.
Behavioral norms in clinical settings vary by jurisdiction, language, and payer system. Generic model fine-tuning loses these distinctions.
The audit trail problem in regulated AI.
Why model providers' audit logs do not satisfy regulator requirements, and what an audit trail engineered for compliance actually contains.
Measuring deflection that matters.
Deflection rate is a vanity metric. What to measure instead: completion-without-escalation, audit-clean rate, clinician override frequency.
Why behavioral context belongs outside the model.
The case for treating behavioral context as a deployable layer, separate from the model. Architectural arguments and three deployment patterns from production work.
47 papers published. 312 citations.
View the full research archive →