Presented at PlatformCon 2025, this talk makes the case that any serious generative AI initiative within an organization deserves a dedicated platform team. Patrick draws on the recurring pattern seen with cloud, mobile, and DevOps adoption: a single team incubates the new technology, a few teams learn to reproduce it, and eventually it must scale across the organization. With AI, the friction point lies between data science teams who have historically owned AI models and engineering teams responsible for production software, with the emerging “AI engineer” role serving as the change agent bridging both worlds.
The talk walks through the three pillars of an AI platform team’s responsibilities: platform, enablement, and governance. On the platform side, Patrick maps out the rapidly expanding infrastructure stack, from LLM providers and vector databases for RAG-as-a-service, to agent infrastructure with memory and state management, MCP server catalogs, execution sandboxes, API gateways for cost and access control, caching layers, prompt tracing and observability tools, toxicity monitoring, and feedback collection systems. He emphasizes that this stack is becoming as complex as the CNCF landscape for Kubernetes, which alone justifies having a dedicated team to manage it.
For enablement, Patrick describes how platform teams should provide prototyping sandboxes, standardized frameworks, golden paths with reusable components and connectors, and critically, support for testing. He highlights the particular challenge of testing non-deterministic AI outputs, where exact assertions fail and teams must rely on techniques like LLM-as-a-judge or human review. He also references his “four AI native dev patterns” framework, noting that developers are shifting from code producers to code managers as AI generates more output, and that platform teams need to be aware of this role transformation when enabling their internal customers.
On governance, the talk covers data protection training, model dependency assessment similar to software bill of materials, emerging AI legislation and risk cataloging, threat modeling, the still-unsolved problem of prompt injection, guardrails for input/output filtering, and PII leak monitoring. Patrick closes with a bonus reference to the unFIX model as an alternative to team topologies, showing how platform concerns exist at multiple organizational layers, not just the technology stack, with his own experience structuring teams across cloud ops, SecOps, developer experience, data platform, and a new AI platform layer.
Watch on YouTube — available on the jedi4ever channel
This summary was generated using AI based on the auto-generated transcript.