Summary
Sandipan Bhaumik, technical lead for Data and AI at Databricks and former AWS principal architect, presents a production deployment playbook drawn from years of enterprise AI engagements at the AI Engineer conference. The talk opens by diagnosing why so many enterprise AI demos fail to reach production: an observability gap (no visibility into model decisions), an evaluation gap (no business-aligned success metric), and a governance gap (no accountability when AI fails at 3 a.m.).
Bhaumik’s answer is a five-pillar framework — evaluation, observability, data foundation, agent reliability, and governance — that teams should design before writing a single line of code. He devotes particular attention to the data foundation pillar, noting it consumes 60% of project time in his experience because enterprise data was built for humans (who forgive errors) not agents (who confidently act on bad data). He distinguishes between “question data” (serving AI outputs) and “tracking data” (tracing, auditing, and online monitoring), arguing both require explicit schema strategy from the start.
The session includes concrete Databricks implementation patterns using Delta Lake, Unity Catalog, MLflow, and Apache Spark, as well as retry/fallback strategies for agent reliability. The talk is aimed at engineering teams and architects preparing to move beyond proof-of-concept AI into regulated, production-grade enterprise deployments.
📺 Source: AI Engineer · Published June 18, 2026
🏷️ Format: Keynote Launch
