DebiAI

DebiAI is an open-source (Apache-2.0) data analysis and bias-detection tool for machine learning projects, developed by IRT SystemX as part of the French Confiance.ai program. It targets data scientists and machine learning engineers who need to examine their data and model results for bias, errors, and contextual weaknesses at the data-analysis and model-evaluation stages of a project. It has been validated in real-world use, including a peer-reviewed paper (ICAS 2024, Best Paper Award) and application to the Valeo WoodScape autonomous-driving dataset, and is deployable via Docker or pip so an organisation's data stays on its own infrastructure. 

The tool connects to a team's existing workflow either through a Python module, which pushes data and model outputs directly from scripts, or through lightweight "Data Provider" web APIs that fetch data live from any source without duplicating it. Data and results are then examined in a customisable, widget-based web dashboard offering bias-detection checks for skewed distributions or labelling inconsistencies, outlier-detection checks for anomalous samples, and contextual model comparison, which breaks performance down by the data slices relevant to the use case rather than a single overall accuracy figure. Users can also select data graphically, via filters and plots, to build new training, cleaned, or investigation subsets, and can extend the built-in checks with custom algorithms through external "Algo-Provider" services. Dashboard layouts and analyses can be saved and re-run automatically as data or models change, with findings exportable to other tools, making bias and error checks a repeatable, shareable process rather than a manual, one-off task.

Auto-discovered on 2026-02-11 by OECD Catalogue Automation