Make the idea playable
At Lobah, I lead a team of four building an AI-powered platform that turns a natural-language conversation into a playable HTML5 game.
The challenge goes beyond generating code. A game needs a coherent concept, a usable interface, matching visual assets, and a working implementation. Every stage needs to stay in sync when the idea changes.
One connected creative workflow
I designed a multi-stage agent pipeline connecting product requirements → wireframes → asset generation → code. A four-branch intent router distinguishes ideation, questions, modifications, and generation so the conversation can move naturally between them.
Different tasks are routed across Gemini, GPT, and Codex. The multimodal workflow combines text, visual wireframes, generated assets, and implementation in one pipeline, bringing end-to-end generation from approximately one hour to fifteen minutes per game.
Built for iteration
- Real-time SSE streaming keeps progress visible during generation.
- PDCA quality loops support refinement across stages.
- A diff-based modification system carries changes through wireframes, assets, and code.
- An S3-backed Game Assets Library stores 21,000+ AI-generated assets, using RAG-based matching to reuse relevant work.
The exhibition photographs are from my personal archive. The process diagram on this page illustrates the workflow; it is not a product screenshot.
