Synthetic data, Training-as-a-Service and visual language models accelerate smart city applications

With new enhancements to its Hafnia developer platform, Danish video software specialist Milestone Systems, in collaboration with NVIDIA, is setting new standards in the development of vision AI applications. The latest advancements were unveiled at NVIDIA GTC in San Jose – with a focus on synthetic data, Training-as-a-Service (TaaS) and Visual Language Models (VLM).

AI training beyond historical data

Traditional AI systems rely primarily on historical data – an approach that quickly reaches its limits in dynamic environments such as smart cities. Rare weather conditions, atypical traffic situations or regional peculiarities are underrepresented in many datasets.

Hafnia addresses this gap by combining real video data with synthetic data generation. Using NVIDIA technologies, realistic, physics-based training data is created that also maps rare or high-risk scenarios. The aim is to reduce biases in datasets and significantly increase the robustness of AI models.

“We enable developers to train models that not only handle familiar situations but are also prepared for unexpected events,” explains Edward Mauser, Director of Hafnia at Milestone Systems.

Training-as-a-Service: Reducing complexity, increasing speed

With its planned Training-as-a-Service offering, Hafnia bridges the gap between data collection and model training. Developers gain direct access to curated, compliant data – both real and synthetic – and can adapt it for specific use cases.

The advantage: instead of fragmented toolchains, an integrated environment is created that simplifies the entire training process. According to Milestone, this significantly shortens development cycles – whilst maintaining high data quality and regulatory compliance.

Visual Language Models for Smart Cities

Another key component is VLM-as-a-Service: a new generation of Visual Language Models that combine visual data with contextual understanding. Based on NVIDIA Cosmos models, these are specifically optimised for urban use cases.

An EU-optimised traffic model is already available and is being deployed in the first cities. Further models are set to follow – for example, for security, mobility or infrastructure scenarios. The aim is to make computer vision applications smarter, more scalable and more cost-effective through generative AI.

End-to-end infrastructure and data sovereignty

Hafnia relies on a multi-cloud architecture that integrates providers such as AWS or Nebius. This allows the entire lifecycle of AI models – from data generation through training to deployment – to be mapped flexibly and scalably.

A key aspect here is data sovereignty: companies retain control over where their data is processed – a crucial factor, particularly in the European context.

Conclusion: From reactive to proactive AI systems

With the further development of Hafnia, Milestone Systems is shifting its focus from reactive AI models towards proactive, resilient systems. The combination of synthetic data, integrated training infrastructure and new model approaches marks an important step towards scalable, practical vision AI – particularly in complex smart city environments.