Show HN: Kapso – Iterative learning from enterprise data
Leeroo's founders, veterans from LinkedIn and Meta, unveil Kapso: an open-source AI engine designed to automate iterative learning and experimentation within complex enterprise data landscapes. This system promises to streamline data work and AI development by continuously learning from deployments, ultimately aiming to outperform traditional methods. Kapso's impressive benchmark results, including leading several public competitions, position it as a significant tool for companies grappling with large, messy datasets.
The Lowdown
Leeroo has introduced Kapso, an open-source AI engine built to tackle the inherently iterative nature of data work in large enterprise environments. Developed by Alireza and Majid, who previously built AI systems for millions of users at LinkedIn and Meta, Kapso aims to automate the process of investigation, hypothesis formation, experimentation, and improvement in data management and AI model deployment.
Key aspects of Kapso and its capabilities include:
- Founding Expertise: The creators bring extensive experience from building large-scale AI systems at companies like LinkedIn and Meta.
- Problem Solved: Addresses the challenge of navigating complex, messy enterprise data estates with numerous source systems, tables, and reports, where traditional, top-down approaches often fail or take too long.
- Iterative Learning Engine: Kapso runs experiments, deploys optimal solutions, and continuously learns from each iteration, akin to reinforcement learning but applied to data and AI workflows.
- Core Mechanism: Features an 'ideation module' that proposes ideas based on past experiments, a 'verifier module' to ensure experiment integrity and evolve success criteria, and a
learn()function to update a 'bank of lessons'. - Git-based Knowledge Base: The bank of lessons is built on Git, enabling sharing of experiences, version control, and continuous improvement.
- Modularity: It is designed to be model-agnostic and coding-agent-agnostic, providing flexibility for integration into various tech stacks (e.g., Docker, Modal, BentoCloud).
- Benchmark Performance: Kapso has achieved leading positions on several public benchmarks, including RelBench for predictive modeling on relational databases, OpenAI's MLE-bench for ML Engineering, and the live International Olympiad in AI 2026, where it outperformed human contestants and most other AI systems.
- Leeroopedia: Offers a public bank for ML and Data with around 100k pages and 6M monthly requests, showcasing its learned knowledge.
Kapso offers a compelling solution for growing companies with intricate data estates and stretched data/AI teams, providing a systematic, AI-native approach to optimize and manage their data and AI initiatives more effectively.