Mahavir Teraiya | Gagan Chawla
Senior Resident Solutions Architect | Director of Forward Deployed Engineering
About
Mahavir Teraiya is a Data and AI Thought Leader at Databricks, helping Fortune 500 enterprises scale agentic AI on governed open lakehouses. With more than 15 years of experience across Databricks, AWS, and Zalando, he designed the first GenAI platform in EMEA with adidas, achieving 91% cost savings. A confirmed speaker at Data + AI Summit 2025 and 2026, and a PhD researcher in federated learning, Mahavir brings a rare combination of production experience and academic depth to the intersection of AI agents, data governance, and open architectures.Gagan Chawla is an award-winning Data and AI leader at Databricks, where he leads forward-deployed engineering and delivery across the DACH region. A business and technical leader with more than 18 years of experience spent entirely in data and AI, he has built and scaled data and AI organizations from the ground up - assembling teams of senior industry experts and guiding them from inception to real, measurable value.
Gagan spends his days helping enterprises move data initiatives and AI agents from flashy demos into governed, production-grade systems, with work that impacts customer outcomes at global scale. He is also a strong advocate for open-source innovation, providing leadership to multiple Databricks Labs projects.
Before Databricks, Gagan held diverse roles at organizations including Mercedes-Benz Research & Development, Teradata, and Infosys. He is a frequent speaker and host on topics related to the realities of operating data and AI at scale.
Talk
Mahavir Teraiya | Gagan Chawla | The Era of Demos Is Over: Engineering Production-Grade AI Agents on Databricks with Unity Catalog, Lakeflow, and Proven Architecture Patterns
AI Agents, Production AI, Databricks, Unity Catalog, Lakeflow, Lakehouse, AI Evaluation
<p>If multi-agent workflows grew by 327% in four months, why are most engineering teams still stuck deploying zero agents to production?</p>
<p>The answer is not more GPUs, better prompts, or bigger models. Databricks’ 2026 State of AI Agents report, based on telemetry from 20,000+ organizations, makes it clear: the bottleneck is engineering discipline.</p>
<p>Databricks’ 2026 State of AI Agents report draws on data from more than 20,000 organizations. According to the report, organizations using AI governance tools move over 12 times more AI projects into production, while organizations using evaluation tools move nearly six times more AI systems into production. Supervisor Agent accounted for 37% of Agent Bricks usage. On Neon - the serverless Postgres technology behind Databricks Lakebase - AI agents create 80% of databases and 97% of database branches.</p>
<p>This is a deep-tech talk for engineers who want the actual architecture, not the pitch deck:</p>
<ul>
<li>The Modern Databricks Stack for Agents (2026 Edition)</li>
<li>Unity Catalog everywhere. Managed tables by default. Lakeflow Declarative Pipelines replacing hand-managed Delta Live Tables (DLT). Streaming tables for real-time agent access. Liquid clustering replacing manual partitioning. Predictive optimization automating maintenance.</li>
<li>Agent-Aware Data Modeling</li>
<li>How AI agents access data differently than dashboards and notebooks. Why a medallion architecture needs an agent-serving layer. How to design schemas that agents can discover, understand, and query safely.</li>
<li>Multi-Format Engineering with Iceberg and Delta</li>
<li>How Unity Catalog’s Iceberg REST Catalog API - read generally available (GA), write in preview - enables agents to query across engines without duplicating data or policies. The engineering trade-offs between Delta managed tables and Iceberg managed tables.</li>
<li>Evaluation Pipelines as Code</li>
<li>How to build repeatable, CI/CD-integrated evaluation pipelines using Mosaic AI and Lakeflow that test agent queries, validate outputs, and enforce cost guardrails before every production deployment.</li>
<li>Failure Modes and War Stories</li>
<li>The $47,000 query. The Cartesian product that crashed a cluster. The agent that leaked PII across dashboards. What broke, why, and the engineering fix.</li>
<li>Code-level detail. Real architectures. Built for engineers who ship.</li>
</ul>
2026-11-25
10:10
10:55