Safe Intelligence

Safe Intelligence develops software and services that help ML developers and validation teams find model fragilities and improve robustness in high-stakes applications. Its platform and command-line tools support formal verification and robust learning for neural networks and decision trees working with vision and tabular data. Spec27, introduced in early access in 2026, extends its offering to specification-driven validation of AI applications and agents. Teams define expected behaviour, generate robustness and security tests, and monitor systems through user-facing access points without privileged model or code access.

Founded in 2021 by Alessio Lomuscio, an Imperial College London professor who serves as CTO, the university spinout is led by CEO Steven Willmott. Its research includes H²V, which combines space-filling dimensionality reduction with Hölder optimisation to assess neural networks’ robustness to rotation, scaling and translation. In its 2025 research account, the company reported validating image-classification models with up to 300 million tunable parameters, including vision transformers; the method’s provable soundness depends on specified conditions.

Safe Intelligence raised £4.15 million in seed funding in 2025, led by Amadeus Capital Partners with participation from OTB Ventures and Vsquared Ventures. It also opened an early-access programme offering tools, model validation and hands-on support for companies in finance, mobility, robotics and aviation, where model errors can carry financial or safety consequences.