BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//sf26-agenda//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
BEGIN:VEVENT
UID:SESS-200@sf26-agenda
DTSTAMP:20261020T151709Z
DTSTART:20261105T213000Z
DTEND:20261105T221500Z
SUMMARY:From Change Data to Live Context: Serving CDC Streams to AI Agents 
 over MCP
LOCATION:Room 2008
DESCRIPTION:Ankit Srivastava\n\nEvery enterprise wants AI agents that act o
 n what's happening right now\, but most agents reason over stale snapshots
  — nightly exports\, cached lookups\, or vector stores rebuilt hours aft
 er the underlying data changed. In a regulated healthcare environment\, th
 at lag isn't just a quality problem: acting on out-of-date member\, claims
 \, or eligibility data carries real compliance and member-impact risk. Thi
 s session walks through how we turned our existing change-data-capture (CD
 C) pipelines into a real-time context layer that AI agents query directly 
 over the Model Context Protocol (MCP) — without adopting a packaged cont
 ext engine. Running self-managed Confluent Platform across GCP and Azure\,
  we use Debezium to capture row-level changes from operational databases\,
  enforce structure with Schema Registry and Avro under backward-compatibil
 ity rules\, and materialize that stream into a queryable serving layer tha
 t always reflects current state. A custom MCP server exposes this layer as
  governed tools and resources\, so agents retrieve fresh\, contextual data
  on demand instead of relying on brittle polling or stale embeddings. We'l
 l go deep on the parts that actually matter in production: topic and schem
 a design that keeps the context layer evolvable\, how we handle PII/PHI th
 rough field-level governance and access scoping at the MCP boundary\, and 
 how we trace and audit every agent query end to end using OpenTelemetry co
 ntext propagated through Kafka headers. Attendees will leave with a concre
 te architectural blueprint for serving streaming data to agents\, a clear 
 view of why CDC is a natural foundation for real-time context\, and a prac
 tical checklist for doing it safely in a regulated domain: schema governan
 ce\, access control\, observability\, and the failure modes to plan for. W
 hether you're building agentic systems or just trying to keep your AI grou
 nded in current reality\, you'll walk away knowing how to make your data t
 he freshest thing your agents touch.
URL:https://sf26.rmoff.net/s/from-change-data-to-live-context-serving-cdc-s
 treams-to-ai-200
END:VEVENT
END:VCALENDAR