A professor couldn't find the right AI tool. So I built a model that isn't allowed to guess.

This project aims to help a friend who teaches economics and commerce find the appropriate AI tools for her research, note-taking, and academic work. When her friend sought assistance for a research paper, she encountered confusion regarding the process. The tool's primary function is to break down the goal into seven distinct tasks and provide corresponding tools for each.

The model never explicitly mentions the tools, preventing the possibility of generating non-existent or outdated tools. The system utilizes a JSON schema to constrain the model's response to the task IDs, ensuring that an invented tool remains impossible. The catalog, a manually crafted array, includes 21 tools that pertain to economics research and teaching.

The front end is a single HTML file, while the server is a small Python script that serves the page and forwards requests to Ollama. The tool remains confined to the user's laptop, ensuring data privacy and eliminating the need for internet connectivity. The setup process involves pulling the Gemma 4 model from Ollama, running a local server, and serving the HTML file.

The tool's limitations include the hand-written catalog covering only 21 tools, which may become outdated, and the lack of ranked tools within categories. The project's cost is minimal, as it runs locally and does not require any licensing fees. The main drawback is the inability to leverage existing tools, such as Excel, and the inability to provide ranked tool recommendations within categories.