Agentic On-Call Platform — for SanDisk
Built an incident-automation system that intercepts PagerDuty events and routes them through a LangGraph intent-and-entity classifier, deciding whether to invoke MCP tools or legacy APIs. The pipeline reduced Sev-1 MTTR by 60%.
Designed a layered agent mesh — Scheduler, Notification, UserProfile — over Postgres and Redis, with GitOps delivery via ArgoCD + Helm. Release stability and deployment speed up 30%.
Notification pipeline uses Redis queues with checkpointed workers and Twilio Studio (SMS + voice) for SLA-grade delivery. Milvus holds embedded SQL templates and metadata for context-aware retrieval; LangGraph composes the reasoning traces.