These weeks were all about wiring up the prompt-handling system. We created the first version of the prompt breakdown logic and established communication between the extension and a Python backend via HTTP, enabling end-to-end functionality.

We added support for image generation via prompts using http://localhost:8000/generateImage. These visuals will help kids better imagine their games. We also implemented the frontend to display these images next to the generated code.

We refined the UI to better suit kids—adding tooltips, simplifying the prompt interface, and making error messages clear and non-technical. A significant advancement during this period was the integration of specialized reasoning and code generation models. We implemented a two-stage AI pipeline where the first model breaks down game concepts into logical components, and the second model generates appropriate code snippets. This approach dramatically improved the quality and educational value of the generated code, ensuring it remained simple enough for children to understand while being functionally correct. We also brainstormed ideas for auditory and visual feedback systems to boost engagement and reinforce learning concepts as children interact with the generated code.

We overhauled our system architecture to support offline usage, a critical feature for educational environments with limited internet access. We setup the QWEN models for offline usage through a Flask app. This resulted in two web servers running in the background: the existing FastSD CPU server for image generation and the new web server for language model inference. We used pyInstaller to bundle these Python scripts into an executable which will be provided to our users along with the model folders, making deployment straightforward for educators. This sprint was also dedicated to testing and polishing. We tested various prompt combinations, fixed bugs like inconsistent AI responses and UI glitches, and resolved issues with prompt parsing. We also gathered peer feedback to uncover issues we hadn't noticed.

Week 21 and 22 were focused around the completion of the project. We first finished the report website. The legal statement was a challenge but we managed to solve it by consulting the client and doing much research. We have also recorded an 8-min video about the project.