Milestone Expands Hafnia AI Platform with Synthetic Data and Training-as-a-Service at NVIDIA GTC
At NVIDIA GTC in San Jose, Milestone Systems unveiled significant upgrades to its Hafnia AI platform, introducing synthetic data capabilities and previewing a new Training-as-a-Service (TaaS) offering aimed at accelerating the development of advanced vision AI models.
The expansion is designed to help developers train artificial intelligence systems beyond traditional real-world datasets, enabling them to handle rare, complex, and previously unseen scenarios. By combining real-world data with synthetic augmentation, Hafnia aims to reduce dataset bias while enhancing the accuracy and resilience of AI-powered applications, particularly in smart city environments.
At the core of this innovation is the NVIDIA Physical AI Data Factory Blueprint, a reference architecture that integrates data curation, augmentation, evaluation, and orchestration at scale. Powered by NVIDIA Cosmos and NVIDIA OSMO, the system converts both real and synthetic data into high-quality, physics-aware training datasets, improving speed and reliability in AI model development.
Bridging the Data Gap
“Together with NVIDIA, we are taking Hafnia to the next level by combining trusted real-world data with synthetic augmentation,” said Edward Mauser, Director of Hafnia at Milestone Systems. “This enables developers to train AI models that are not only accurate in known situations, but also resilient in the unexpected.”
Traditional AI systems rely heavily on historical data, which often lacks representation of rare events such as extreme weather, unusual traffic patterns, or region-specific conditions. Hafnia addresses this challenge by integrating synthetic data into its curated video library, allowing developers to simulate dangerous or uncommon scenarios while maintaining compliance and data authenticity.
Training-as-a-Service Simplifies AI Development
Milestone also introduced a preview of its upcoming TaaS offering, which provides developers with streamlined access to Hafnia’s compliant video datasets—both real and synthetic—within a unified training environment. The service eliminates the need for fragmented data pipelines, allowing developers to focus on building high-performance video analytics solutions.
With customizable datasets and fully traceable data sources, the platform ensures regulatory compliance while significantly reducing development complexity. According to Milestone, the solution can accelerate AI model training processes by up to 30 times.
VLM-as-a-Service for Smart Cities
In collaboration with NVIDIA, Hafnia also offers VLM-as-a-Service, featuring Visual Language Models optimized for smart city applications. Built on advanced reasoning models, the platform is already being deployed in select European cities for traffic management.
Milestone reported notable performance gains in these models, including a 19.4% improvement in traffic flow accuracy, an 8.9% boost in visual detection capabilities, and a 4.4% increase in alert verification accuracy. The company plans to expand its VLM offerings to cover a wider range of urban scenarios.
Scalable Multi-Cloud Infrastructure
To support end-to-end AI development, Hafnia operates on a flexible multi-cloud framework leveraging providers such as Amazon Web Services and Nebius. This approach ensures scalability while addressing data sovereignty requirements, allowing organizations to maintain control over sensitive data.
Additionally, Milestone announced the launch of its synthetic data generation pipeline on Nebius, further enhancing its ability to support the full lifecycle of vision AI development—from data sourcing and training to deployment.
With these advancements, Milestone Systems positions Hafnia as a comprehensive AI development ecosystem, enabling faster, more reliable, and scalable deployment of next-generation vision AI solutions.
