Zer0Fit - Zero-shot predictions, classifications, and regressions using Google ML research models running locally as a dockerized MCP

AI grad student here. With all the recent interest in Jev, I thought I would share something I built la few months ago that brings ML models and LLMs together in a different way than Jev does for different use cases. github.com/porespellar/Zer0Fit Background: A few months ago on their research blog ( research.google/blog/introducing-tabfm-...shot-foundation-model-for-tabular-data ) Google released TabFM zero-shot foundation model for tabular data. It was kind of ignored except by maybe a few machine learning nerds that care about that kind of thing. I mean, for real tho, TabFM wasn’t exactly the sexiest name choice. I personally thought TabFM was cool as shit because it kind of melded classical machine learning models into an LLM of sorts. So anyways, I wrapped Google TabFM (their model for classifications and regressions), and Google TimesFM (their model for predictions) into a convenient Fast API and made the whole thing a dockerized MCP that you can connect to your favorite LLM. I call my project Zer0Fit - Zero-shot ML tasks without needing to train or fit a model. Here’s my repo if you want to check it out: github.com/porespellar/Zer0Fit You see what I did there with the name? It took me hours to come up with that name :) I’ve made it as easy as I could to install. Just clone it and run the install script. So the basic idea is, you connect the MCP to whatever LKM you want, give it a dataset (CSV, tabbed data, or time series), and ask it what you want it do do with the data. it decides which of the Google models to use, and then it runs the regression, classification, or prediction task in context and gives the results back to your LLM. That’s the best way I can describe it. See the Google blog for the details on what the Google models are actually doing. Again, I’m not doing anything special, I’m just wrapping the Google models up to serve locally and making them exposed via MCP. The Google models are doing all the heavy lifting. I have absolutely no connection to Google research and am not associated with them in any way other than being a fan of them releasing this for us to try locally. Is it better than a data scientist building a custom model to do an ML task? No, definitely not, but it is much easier, and probably will get you an answer that is reasonably close (or possibly at least in the ballpark) and that might be good enough for some use cases depending on what you’re looking for (assuming it’s not a task that requires high precision, or high speed classification). Anyways, I just thought the Google models deserved some attention and love from the community, so I wanted to make them more accessible, that’s all, that’s why I made Zer0Fit. If you want to try it out it’s over on my GitHub in the link above. Please remember, this is all just stuff. It’s cool to play with, but don’t use this with anything where it’s output matters. Use at your own risk. P.S. I made it with Open WebUI in mind so it should work well in that, but it’s an MCP so it should work with just about anything that is MCP-friendly.