Safe Sign Technologies

Legal AI lawyers can trust.

High Quality Data at Each Training Stage
Safe Sign's AI systems are built with a data-centric approach, with data drawn from proprietary data sources and carefully reviewed by their legal team, constituting the highest quality legal dataset anywhere in AI. They use new curriculum learning algorithms during training, drastically reducing the time and cost of pre-training. Subsequently, they analyze the model using active learning and adversarial methods, ensuring additional data collection is guaranteed to improve safety & performance.

General Purpose Legal Foundation Models
Their models are trained on the foundational properties of legal reasoning, independent of jurisdiction or particular legislation. This allows them to build systems that are customized to a wide range of legal tasks, enabling them to solve problems spanning contract negotiation, legal Q&A, dispute resolution, T&C product rankings, and many more using a single model.

Safety, Reliability & Robustness
Safe Sign emphasizes safety throughout the development process with adversarial testing and red-teaming carried out by legal experts. Their models feature out-of-distribution detection, uncertainty estimation, and agent-based roundtable discussions to provide accurate and explainable responses. They specialize their system to specific jurisdictions and drastically reduce hallucinations using Retrieval Augmentation (RAG), requiring citations to up-to-date legislation with every model response.

Model Efficiency & Real-time Inference
Their team has world-leading experts in sparsity, model distillation, and quantization, ensuring that their predictive models run at real-time speed and can serve thousands of requests.

Safe Sign was acquired by Thomson Reuters in 2024.