AI
AI-Generated Terraform from Natural Language Intent
Application Design Center takes a natural-language description of a use case and produces Terraform configured to Google Cloud best practices. Engineers who know their use case but not Terraform can reach 80-90% of a working solution without touching infrastructure code directly. Platform teams review, approve, and promote the result to a shared catalog. One attendee at the previous panel noted that environment creation used to take four weeks. With AI-generated, pre-approved templates, that drops to hours.
Grounding AI on Private Infrastructure Patterns
AI models trained on public GitHub code don’t know your company’s internal patterns. Application Design Center solves this by accepting a pointer to a private Terraform repository. From that point, anything the tool generates is constrained to company-approved components, or a configurable mix of internal and Google-provided modules. Platform engineers can mute out Google defaults entirely and allow only their own. The system treats private repo content as vetted ingredients, so developers get self-service access without the ability to introduce unapproved infrastructure.
Secure and Compliant by Design
The compliance model has two modes. For HIPAA, PCI-DSS, or FedRAMP environments, platform engineering and the CISO office define golden templates as the only self-service options. Anything outside those templates takes four weeks of manual review. For teams that need more flexibility, the middle path is pre-deployment linting: OPA rules, Sentinel policies, Wiz, or Security Command Center all run against the design before it reaches runtime. Changes are cheap at design time. After deployment, they are not.
Observability and Cost Awareness Built Into Infrastructure
Telemetry baked in at provisioning time means AI can query logs against a named application the moment it goes live. Manasra pointed to the Gemini Investigations root-cause demo (21:04): 502 errors traced to a misconfigured node pool, with a billing cap flagged as the blocker. The same logic applies to cost. An AI agent running on a $30,000-per-month Nvidia G4 cluster and one on a pay-as-you-go managed endpoint are both valid choices, but the right answer depends on SLO targets, not guesswork.
Trusting AI at 80 Percent
“trust it to 80%. trust it as an accelerator” — Wael Manasra
That is the practical posture Manasra recommends: AI drafts the Terraform, humans do the code review and decide whether to deploy. Four weeks shrinks to hours plus a human check.
“That to me is success. That is the cheat code, you know?” — Wael Manasra
The risk companies fear is losing control. His answer: the CI/CD review gate stays in place. AI fills the gap between intent and working infrastructure at 80-90% of the way there, then hands off. Engineers own the final decision.
Notable Quotes
trust it to 80%. trust it as an accelerator Wael Manasra · ▶ 27:42
but getting you to 80 or 90% faster is just such a huge productivity boost Wael Manasra · ▶ 28:15
That to me is success. That is the cheat code, you know? Wael Manasra · ▶ 28:22
Key Takeaways
- Application Design Center generates best-practice Terraform from natural language, reducing environment setup from four weeks to hours.
- Enterprises can ground the AI on private Terraform repos so all generated code uses company-approved, pre-vetted components.
- Treat AI as an 80-90% accelerator with human review at the end, not an all-or-nothing replacement for engineering judgment.