Web Gen

Experience the future of Vibe Coding. Web Gen uses Gemini 3 Flash to transform natural language into complex, responsive web interfaces, handling everything from logic to UI in one click.

Comment

Inspiration

Web development is often a battle between creativity and syntax. We were inspired by the concept of "Vibe Coding"—the ability to build complex, functional software through natural language alone. We wanted to eliminate the "blank page" problem for developers by creating an agentic partner that doesn't just suggest code, but architects entire web experiences in seconds.

What it does

Web Gen is an AI-powered IDE that transforms high-level ideas into fully self-contained, production-ready web applications. By entering a simple description, users receive a single-file deployment (HTML/CSS/JS) that is responsive, interactive, and ready for use. From landing pages to Canvas-based games and data dashboards, Web Gen handles the logic, styling, and structure autonomously.

How we built it

The core engine is built on the Gemini 3 Flash model. We chose Flash for its industry-leading low latency, which is essential for "interactive coding" workflows. Agentic Reasoning: We utilized the thinking_level: high parameter, allowing the model to reason through the DOM structure and CSS grid logic before generating code. Thought Signatures: To allow for iterative design (e.g., "now make it dark mode"), we persist Thought Signatures across turns, ensuring the AI maintains the architectural "context" of the original build. Mobile-First IDE: Built with a custom Android framework that integrates AI Studio’s API for near-instant execution.

Challenges we ran into

Maintaining contextual state in single-file generations was difficult. Standard LLMs often forget the CSS variables they defined earlier in the prompt. We solved this by implementing a multi-stage agentic loop where Gemini first "plans" the global styles and then "executes" the component logic, ensuring a cohesive and bug-free final file.

Accomplishments that we're proud of

We achieved a "Zero-Error" generation rate for complex interactive components like Tic-Tac-Toe games and multi-tab dashboards. We are also proud of the Auto-Save & Project Library system, which allows users to build up a modular library of AI-generated components directly on their mobile devices.

What we learned

We discovered that Gemini 3's reasoning is significantly more "spatial" than previous versions. It understands UI layouts (like "centering a div" or "responsive sidebars") much better when treated as an architect rather than a simple text-completer. We also learned that low-latency AI response is the key to keeping users in the "creative flow."

What's next for Web Gen

Our next step is Multimodal UI-to-Code. We plan to use Gemini 3's vision capabilities to allow users to draw a website on a napkin, take a photo, and have Web Gen turn it into a live website instantly. We also aim to add Firebase hosting integration so users can deploy their generated apps to the web with one click.

Share this project:

Updates