KaggleGPT

Inspiration

I wanted to make this project because I am an aspiring Kaggler entering university this year but being new to the scene I find data science and model building quite intimidating so I thought I'll create something to help myself. I also wanted to see how far the latest AI models can go in competing in these competitions end to end against humans.

What it does

KaggleGPT takes the csv input data files provided as Kaggle competitions and runs it through the core agentic workflow. So there will be agents that handle: exploratory data analysis cross-validation methods evaluation metric generations feature engineering feature selection model building critic agent the models will be made in an experimental loop, recorded in an ablation table.

How we built it

I built it mainly using Chatgpt for the ideation stage and codex for basically the entire workflow.

Challenges we ran into

The main challenge I ran into was a constraint on tokens and compute. I ran out of tokens even after using 4 weekly resets I had saved up + $21 in API credits. I had to find more cost-friendly ways to prompt and build the system.

Accomplishments that we're proud of

A roughly working model that proves the future of ML engineering for Kaggle comps will be made easier using AI.

What we learned

I learned that while mimicking the general workflow of a Kaggle competition is relatively simple because there exists a repeatable structure towards approaching each of the competitions, it is difficult to properly create a system that can generalise and grasp the nuances between different types of competition. I tried solving that nuance and ambiguity through employing an agentic system but to produce usable non-generic structured outputs required multiple high functioning models which are costly at scale and might be a struggle to justify for a Kaggle competition.

What's next for KaggleGPT

What is next for KaggleGPT is to further strengthen and work on the various components of its agentic system, to separate LLM inference needs from specific Kaggle workflows that can be done through deterministic code. I think that will reduce the costs needed to run this system and make it more viable in the future.

Built With

• nextjs
• typescript