Xpdeep Information

Most industrial or critical AI never reaches production. Not because the models underperform, but because nobody can certify, insure, or defend them. Xpdeep removes that constraint at the architecture level. We build deep learning that is natively explainable. The explanation is a structural property of the model, produced during training, not an approximation reconstructed afterwards by SHAP or LIME. On multivariate time-series, accuracy is at or above black-box equivalents: joint optimization of precision and structural explanation eliminates wasted capacity. Governance. Risk quantifies exposure. Legal defends the decision. Compliance produces EU AI Act and ISO 42001 documentation as a by-product of the model lifecycle. Insurers underwrite. Certifiable by design, or not certifiable at all. Alignment. Models steer toward the real KPI (margin, yield, energy cost, asset availability) instead of a proxy. Non-contributing variables are removed before deployment: on one industrial benchmark, 105 inputs down to 19, running 35x faster. Prescription. What no black-box architecture can reach. Every inference returns the prediction with its structural explanation, then the minimal input change required to reach a different outcome, then why that change is the minimal one. Minimal causal intervention, structurally justified. A steering system, not a reporting system. The platform. XpViz for data science teams. XpAct for operations and business users. XpComply for legal, risk, and compliance. Native mode trains within the Xpdeep architecture and unlocks the full chain through to prescription. Retrofit mode converts existing PyTorch models for understanding and audit, as an on-ramp. Where it runs. Automotive, aerospace and defense, energy and nuclear, semiconductors, industrial systems. SaaS, on-premise, or sovereign cloud, including air-gapped deployments. Spun out of research at the Laboratoire d'Informatique de Grenoble (Université Grenoble Alpes / CNRS).