Hey HN! We're Alireza and Majid building Leeroo (https://leeroo.com/). Previously we’ve built AI systems for products with millions of users, including at LinkedIn and Meta. We’re now focusing on building AI systems for enterprise data. Today, we're open-sourcing Kapso (https://github.com/Leeroo-AI/kapso), an engine that runs experiments, deploys the best solution, and learns from every run.The nature of data work is iterative: investigate, form hypotheses, experiment, and improve based on what you learn. This bottom-up process is the only way to deliver meaningful outcomes quickly in large, messy enterprise data environments. Consultants gathering context in meetings or teams defining a business ontology upfront to fit into the data has no chance of execution, in quarters and sometimes years. An AI-native system should automate this process, compounding learning to each new investigation and implementation.For instance, one of our enterprise projects was moving a company off its old data estate onto a lakehouse with more than 30 source-system families, thousands of tables and transformations, and hundreds of BI reports. No individual or AI agent knew how that data estate was connected and operated. The only way to learn it was to investigate the company's repos, dbt models, documentation, and data itself and propose hypotheses about the estate, run experiments, and learn from them, until we understand how each system behaves and under which conditions.The iterative work goes beyond understanding how data and systems fit together, and the business operations. Building useful software is iterative too. Whether you’re optimizing a warehouse, building a lakehouse pipeline or predicting churn, you need to run experiments and learn what works best in that environment. As you deliver more projects, that know-how compounds, similar in principle to RL. RAGing documentation has little value until it has been tried in a specific environment and the consequences are learned. As with human engineering work, experimentation and learning should happen in a sandbox.We built Kapso for this kind of work, you give it an objective and success criteria, and it works towards an optimum solution that can be deployed on your stack.You can view a video walkthrough here: https://www.youtube.com/watch?v=W4FHlza1fRsHere’s how it works…Kapso has a bank of lessons that are induced or deduced from the past experiments, and each carries a credence score, conditions on where it applies, and the evidence behind it. For each new experiment, an ideation module proposes ideas considering the earlier lessons, re-ranks them by expected return against the budget each would use, and sends the top-ranked ones to the implementor. Because the bank keeps learning from the past experiments in that