Google engineer demos 'vibe coding' tools for Hawaii students; emphasizes planning, security and version control

Divya Soni, a Google customer engineer who supports the state of Hawaii and higher education, led an online workshop demonstrating “vibe coding” — using generative AI to create code from natural-language prompts — and walked students through tools including Gemini, Firebase Studio, Stitch and AI Studio.

Soni said vibe coding lets people convert ideas into working prototypes in minutes, but cautioned that prototypes are not production systems and that developers must test and secure any deployment. “Natural language is your new code, is your new programming language, if you will,” she said during the session.

The session combined a live demo and step‑by‑step guidance. Soni showed Gemini generating an initial website prototype for a sample dog‑walking app called “Aloha Paws,” then moved the same idea into Firebase Studio to link the prototype to Google Cloud services and to deploy a preview. She also demonstrated Stitch for UI design and AI Studio for more controlled model‑driven authoring. Participants asked about language choices, testing, and security; Soni replied that prompts can specify languages such as Python or Go and that AI can assist with testing but users must verify results.

Why the workshop matters: Students preparing for an upcoming hackathon were the target audience; the tools Soni demonstrated can cut ideation‑to‑prototype time from days or weeks to minutes but come with operational and governance requirements. She emphasized several practical rules for using AI safely and productively: develop a clear plan and minimal product requirements before prompting an LLM, iterate and debug code, use version control (Git/GitHub), avoid entering personal identifiable information (PII) into public LLM interfaces, and perform security reviews before production deployment.

Soni explained that some Google tools used in the demo are free to students in their web form (for example, gemini.google.com and AI Studio), but deploying prototypes to Google Cloud Platform (GCP) requires a GCP account. She said students typically receive $300 in GCP credits for learning and testing. On security she described it as a shared responsibility: the models will produce “skeleton” code or placeholders for authentication and authorization, but developers must configure and connect those pieces to their own identity and billing infrastructure.

On technical details, Soni walked through prompts that request starter code, then iteratively asked Gemini to add features (for example, a booking feature). She noted that the first solution produced by an LLM will often require refinement and debugging. She also explained model choices available through AI Studio — for example, Gemini 2.5 and smaller “flash” variants for faster, lower‑cost inference — and advised teams to pick a single, appropriate model for an application rather than mixing models.

Participants put forward real questions during the demo. Jessica Wood asked whether Gemini can update an already active website; Soni showed Gemini producing an HTML skeleton and said developers can prototype and then rebuild or replace site components but must perform testing and security work before publishing. Another participant asked about input validation and security controls; Soni reiterated that configuration and deployment of authentication/authorization are up to the developer.

Soni closed by encouraging attendees to experiment, start with small, testable features, and use version control to protect work. She said she will share the slide deck, links to the tools shown and a recording to the event Slack channel and other course materials.

The workshop was presented as a skills and tools briefing rather than a policy or procurement action; no formal decisions or votes occurred during the session.