Sanjay Ashok

As organisations race to deploy AI across the enterprise, many are discovering that the hardest part is not choosing the right model. It is giving AI enough business context to produce accurate, reliable, and cost-effective results at enterprise scale. This panel explores what "context" actually means for enterprise AI, where semantic layers, ontologies, and governance fit into the picture, and what organisations are learning as they move AI initiatives from pilot to production. Drawing on real-world examples, these industry leaders will share practical lessons, common pitfalls, and the context that makes the difference between AI and enterprise AI.

CMSPI built a Databricks-native analytics app that automates payments optimization at enterprise scale — replacing spreadsheets with a governed platform that processes billions of transactions and delivers measurable time and cost savings for global merchants. Lakebase powers real-time scenario configuration, while Delta tables and SQL warehouses enable interactive analytics. Databricks Workflows orchestrate a mathematical solver (PROM) that evaluates billions of routing combinations and returns results in 30–40 minutes — eliminating 30+ hours of manual work per initial analysis and uncovering ~10% incremental savings opportunities. This session shares CMSPI’s path to productization: access control with Entra ID groups, MLOps for approval-rate prediction models and external monetization via Delta Sharing to non-Databricks consumers. The architecture demonstrates the “built on Databricks” pattern for secure, governed apps — with plans to expand collaboration using Databricks Clean Rooms.

Mobile game studios run multiple experiments simultaneously, but daily insights are often delayed by manual data preparation, inconsistent methods and ad hoc notebooks. This presentation shows how SEGA HARDlight migrated from Tableau to Databricks and built an automated framework for experiment analysis. It ingests definitions and telemetry, applies uniform statistical models and publishes a daily dashboard with an LLM-generated summary and KPI diagnostics. Unity Catalog manages assets and lineage, while Lakeflow streamlines pipelines. MLflow ensures consistent experiment tracking, and AI/BI dashboards with Genie enable natural language exploration. The dashboard locks in results and decisions, creating an auditable record of truth. Early estimates show over eight hours saved weekly for the data team and roughly half a day saved per new query through automation, aiming to double testing capacity — while also boosting trust and consistency across experiments.

At SEGA Europe, we process billions of game telemetry events daily to power real-time analytics for live ops, player experience optimisation, and business intelligence. We have rebuilt our streaming platform on Lakeflow Spark Declarative Pipelines, moving from fragile, manually managed jobs to a resilient, self-healing lakehouse architecture. This approach was battle-tested during the launch of Football Manager 2026, where scale and reliability were critical. This session shares our production journey with Lakeflow—the wins, lessons, and unexpected challenges of operating at gaming scale. We’ll cover our migration from legacy streaming infrastructure, achieving sub-minute latency, and handling evolving event schemas. Topics include stream fan-out from Kinesis to Bronze, schema evolution, late and out-of-order events, data quality checks, performance tuning, and cost control. We’ll also share quantified results from FM26 and practical patterns for teams evaluating Lakeflow at scale.