since several tasks in recsys part have text-based columns, the agent , after experimentation, decided to use specialized LLM and/or encoder ( e.g. MedCPT encoders) to transform those columns to effective features for the prediction.
Alireza Mohammadshahi
AI & ML interests
Recent Activity
Organizations
@jprzybylko @valterh-nv @flopeznv @akihironitta-nv @RBendiasN @aw777 @profjure @rusty1s
Hi Team,
I’m Alireza, co-founder of Leeroo (YC-backed). We built Kapso, a harness for data engineering and ML experimentation.
Kapso outperforms KumoRelational on the official RelBench leaderboard: 81.2 vs 79.3 AUROC (higher better) and 24.8 vs 28.5 nMAE (lower better). Your own October 9 submission report already shows these results, predating this post. The claim about outperforming the strongest supervised models needs to be updated.
Our hypothesis: building better harnesses offers faster iteration and more flexibility than developing a new relational foundation model. A harness can adapt the data pipeline, model choice and objective without another foundation-model pretraining cycle. RelBench illustrates that breadth: Kapso leads classification, regression and recommendation, while KumoRelational’s cited submission covers only classification and regression.
That makes harness engineering a concrete winner vs. building an expensive foundational table model.
Links:
Kapso Github: https://github.com/Leeroo-AI/kapso
Official Relbench link: https://star-project.stanford.edu/relbench/leaderboard/#classification
DABStep: Data Agent Benchmark for Multi-step Reasoning


- +5


