cactus-compute/needle ? reverse-engineered prompt

Build me a Python project for a tiny local AI model that takes a normal user request and chooses the right function call from a list of available tools. I want it to feel like a lightweight personal AI helper that can run on small consumer devices, not a big chat bot.

Include a simple web playground where I can type a request, paste in tool descriptions, see the predicted tool call, and finetune the model on my own examples. The weights should download automatically the first time so setup feels easy. Add a clean Python API for loading the model, running a query, and getting back a JSON style tool call. Also add command line commands for playground, run, finetune, train, evaluate, tokenize, and synthetic data generation.

Please include clear example data for weather and smart home tools, plus instructions for creating enough training examples per tool so the model does not just memorize. Keep the README practical so I can clone it, run setup, open the playground, and test it quickly.