T H E W A L K

Design Technologist,
AI Infrastructure,
Prototyping.

← Back to main

Deep Dive: CI/CD & Deployment Infrastructure

Focus: Building deployment infrastructure as a one-person shop inside a high-compliance environment

1. The problem: infrastructure friction

At Google, moving from a research insight to a deployed AI prototype often takes weeks. The bottleneck isn’t model capability — it’s fragmented tooling, complex networking (UberProxy/VPC-SC), and boilerplate. I needed a deployment pipeline that enabled rapid iteration while maintaining security standards.

2. What I built

A. The deployment pipeline

The core challenge was bridging internal GitHub Enterprise (Depot) with Cloud Run across restricted VPC boundaries.

  • Layer 4/7 hybrid routing: Custom routing using an Internal TCP Proxy Load Balancer (L4) to wrap Private Service Connect (PSC) connections. This made the internal Git repo addressable via a static internal IP, bypassing standard proxy blocks.
  • Cloud Build orchestration: Non-standard source-fetching via Developer Connect Git Proxy, since standard CI/CD tools couldn’t natively handle the restricted network.
  • Zero-trust ingress: Every prototype deployed to Cloud Run behind an Internal Application Load Balancer (L7) with Identity-Aware Proxy (IAP) — accessible to anyone at Google, invisible to the public internet.

B. Preview environments

To accelerate prototype iteration, I built pre-configured templates that abstracted away infrastructure boilerplate.

  • Dash-preview (ephemeral environments): Any branch with a *-preview suffix automatically provisions a unique, live URL. Researchers and designers can test UI changes in seconds.
  • AI-native code review: Integrated a Gemini-powered reviewer into the PR flow for instant feedback on security and performance.

C. Results

  • 5-10x deployment frequency: Up from 1-2 times per day. The Agent Builder team deploys 5-10x daily on this infrastructure.
  • 48x research efficiency: This pipeline powers the UXR Platform, reducing per-study analysis from roughly 4 hours to 5 minutes.

3. Scale

50+ direct UX/PM/Engineering adopters and multiple teams, with a broader 200+ stakeholder surface; 3,000+ commits and 1,500+ PRs across the cloud-ai-ux organization. Multiple Google orgs forked the structure. A single engineer building the right infrastructure moved an entire team from friction to flow.

← Back to main