Gemini Enterprise: Keynote Demos & the Tools Behind Them
Shipping a keynote in a week
Cloud Next ‘25 came down with an aggressive timeline — a working Conversational Commerce demo, on stage, in front of customers and press. I shipped it in under a week. That speed wasn’t heroic hours; it was the tools I’d already built for my own velocity:
- Figma MCP Server — a Python MCP server exposing the Figma API to AI coding agents. Cross-file component resolution, simplified YAML output, targeted node fetching. Months before Figma shipped their official MCP.
- Material Components library — a clean component library generated by exporting from Figma through the MCP server. Tokens and components stayed in sync with Material’s source of truth, so the AI agent had a real component vocabulary to draw from instead of inventing CSS.
- Early AI coding agents — pairing the MCP server and the component library with an agent that could read the Figma file, look up the right component, and generate working UI in one pass.
Combine these and “here’s the design” turns into “here’s the deployed prototype” in hours instead of days. People at Google noticed. The team started asking how I was shipping that fast — which is the moment a personal velocity recipe started turning into something I had to make repeatable for others.
Conversational Commerce (Cloud Next ‘25 Keynote)
The highest-visibility thing I built at Google. A conversational shopping app where users type natural language queries (“show me red shoes for a summer wedding”) and get AI-generated responses with product carousels, suggestion chips, and follow-up refinements — all within a flowing multi-turn conversation. Peer Bonus for keynote delivery.
Stack. Vue 3 + TypeScript + Pinia on the frontend, with a single orchestrator tracking the conversation lifecycle across turns. Two Python Cloud Functions on the backend wrap Google’s pre-release Retail v2alpha API — one for streaming conversational search (text, follow-ups, refined queries), one for product detail (variants, fulfillment, pricing).
Key decisions:
- Polymorphic message stream — text, carousels, and chips are all
ChatMessageobjects rendered by type. Flat stream, not separate state per element. - Concurrent refined searches via
Promise.allSettled— one failing query doesn’t block the others. - Mobile-first: touch scrolling, sticky input above virtual keyboard, responsive carousels.
The broader architecture. Alongside the working prototype, I designed the Root Orchestration Agent architecture on ADK — a hierarchical system where a central agent decomposes user goals into tasks delegated to Specialist Agents (Commerce Search, Scheduler, Coupons, Compliance).
Conversational Commerce — enterprise agent orchestration
Making the recipe repeatable: the team environment
Once others wanted in, the velocity recipe needed to become repeatable. The Agent Builder team was building prototypes in Google AI Studio — each one an isolated experiment, no version control, no shared components, no way to deploy for user studies. I used the GitHub Enterprise platform I’d set up for Cloud UXE to give the team a proper home, and built the embedded tooling that turned a clean checkout into a working prototype environment.
Repository migration and onboarding. Moved AI Studio prototypes into a structured GHE repo powered by Jetski. Wrote the onboarding guide and personally taught designers how to branch, commit, and open PRs.
Automated CI/CD. Cloud Build pipeline where merges to main automatically build a container and deploy to Cloud Run.
Preview branches. Suffixing any branch with -preview auto-generates a unique, isolated URL. UX Researchers get stable links for user studies; stakeholders get “frozen” demo URLs physically separated from development.
Automatic preview branch deployment with isolated sandbox URLs
Embedded prototype tooling:
pr-reviewskill — pushes changes and opens a PR with a diff summary comment.- Visual Inspector — color/typography tokens, live property editing, Material standards comparison.
- Annotate (Draw) — draw over the prototype, screenshots auto-saved to clipboard.
- Commenting — Figma-style Firebase-backed comments with @mentions.
Visual Inspector — real-time Material token auditing
The environment caught the attention of CoreUX, Corp Eng, and Google Material. That demand is what led to the GitHub Enterprise Platform & vibe CLI — the org-wide generalization of this same workflow, for any team.
Marketplace Demo (Cloud Next ‘26 Partner Keynote)
Brought in for rapid execution of a multi-agent ecosystem demo — specifically requested because of the reputation the ‘25 keynote built for shipping fast. Delivered a working scaffold within days of first contact. 6-8 minute live demo showing autonomous agents handling enterprise procurement — vendor selection, compliance validation, logistics orchestration. Same velocity recipe, now running on the matured team environment instead of a personal setup. Peer Bonus.
Cloud Next '26 Marketplace Demo — multi-agent orchestration