Infrastructure, Observability & Security
US healthcare clients need controls that survive audits without slowing every release. Monitoring must prove who saw what, what the model used, and when humans overrode it. We deploy in regions you approve, with separate accounts for prod and non-prod. Secrets never live in prompts or tickets. Network policies deny broad outbound internet from inference workers by default.
Compliance programs cover HIPAA safeguards and align to SOC2 style controls when buyers require vendor evidence. BAAs, encryption standards, and access reviews run on a fixed calendar. For Virginia providers with federal limited scopes, we document boundary devices and log shipping destinations. Least privilege maps from group membership, not shared service accounts that linger after contractors leave.
What we monitor includes p50/p95 latency, token spend, refusal rates, retrieval miss rates, and human override reasons. Drift checks compare input length, specialty mix, and class distributions week over week. Cost alerts fire before invoices settle. Queue depth alarms protect staff from silent pileups on Monday mornings. Synthetic probes hit critical paths every few minutes.
Incident response names severities. Suspected PHI exposure pages security and legal within minutes. Model quality drops page product and clinical owners with sample packs. Runbooks list containment steps: freeze writes, replay last known good version, and open patient outreach only when instructed by clinical leads. Postmortems mark whether process or design failed, then update goldens and gates.
Deploys for Alexandria and broader US clients prefer blue-green with traffic mirrors. Infrastructure as code keeps drift visible. Backup and restore drills prove index rebuild times before you need them. We teach your team the dashboards so exit is real. Vendors that keep the only keys fail this bar and do not belong outside temporary pilots.