The lab had a requirement to publicly offer dataset downloads while also tracking how many people have downloaded the data, leading to this project.

CITE Lab MoCap and Humanoid Group Download Center

Initially, I planned to use FastAPI + SQLite—my preferred stack—and fed the following prompt directly into AI (Claude 3.7 Sonnet), which produced a surprisingly usable result. However, two bugs emerged: one was that SMTP email sending wasn’t using encryption, and the other was that NJU Box used an outdated version, requiring manual lookup of the corresponding Seafile API.

However, the resulting interface was mediocre, and many features were hard to scale. So I decided to refactor the project using more advanced tools like Next.js. But whether I asked the AI to refactor existing code or start from scratch with prompts, the results were poor—often including compilation errors rather than runtime issues. I also tried several latest AI IDEs, but they didn’t deliver satisfactory outcomes.

In the end, human effort was unavoidable. I spent a full day manually setting up the framework, then gradually fed each file from the original FastAPI implementation into the AI, asking it to reference and complete the corresponding functionality in the current Next.js project. Then, step by step, I guided it to improve small features, fixing bugs manually along the way, until finally achieving the result seen in the two images above.

Current Features:

• Tech Stack: Next.js + TypeScript + Tailwind CSS + Ant Design + Seafile API + MongoDB + Redis
• Frontend provides a download interface where users can select files for download (supports multiple selections).
• Supports downloads only for academic or educational emails, requiring valid email verification before download.
• Includes download terms that users must read before proceeding.
• Backend logs download records including username, filename, download time, etc., and provides an admin dashboard for viewing.
• Files are stored on Seafile (NJU Box). This project generates only temporary links for users to download, avoiding direct exposure of file URLs and saving server bandwidth.
• Parses the readme.md file within Seafile and displays it on the download interface as file documentation.
• Supports configurable subpaths for deployment within existing websites.
• Supports dark mode.
• Uses AI to parse organization names based on email domains.

@Que: "There's no bad AI—only humans who aren't stubborn enough."