Milestone reveals city AI model as hackathon prize hits EUR €5,000
Milestone Systems has showcased its upcoming Vision Language Model (VLM) to artificial intelligence developers from around the globe in a hackathon designed to enhance video intelligence for urban environments. Fifty-five developers were given early access to the VLM, trained on more than 75,000 hours of curated video data, with finalists announced following a competitive event.
Product expansion
The VLM, built using NVIDIA Cosmos-Reason and Hafnia's domain-specific data library, is designed to interpret city visuals, languages, symbols, and events. Milestone is also introducing a generative AI plugin for its XProtect Video Management Software, with an emphasis on improving traffic management in complex urban settings. The plugin turns video feed into actionable, detailed written reports, summaries, and live alerts.
Hackathon challenge
Participants from 15 countries were tasked with leveraging the VLM via API to create integrations with third-party applications, targeting smart city improvements. The brief focused on solutions that could make use of video streams in a more efficient and privacy-aware manner. Currently, analysts face the challenge of extracting meaningful insights from vast volumes of video footage-a process Milestone aims to address through automation and artificial intelligence.
Winning solution
"I wanted to build something that is relevant and useful in real life. What I liked the most about the hackathon was how easy it was to get started. The API and documentation made it simple to build a demo quickly, brainstorm more ideas, and try them out," said Thomas Kreutz, winner of the hackathon.
Kreutz's entry, Ask The City, uses Hafnia's VLM API to convert live city camera footage into instant, privacy-conscious responses to user queries. Users select locations on a map and ask questions in natural language, with the system returning answers generated from the most current video frames. Kreutz received EUR €5,000 and a NVIDIA Jetson AGX Thor Developer Kit.
"Technology is evolving faster than ever, and no single company has all the answers. The winners are those who can bringtogether the best technologies, the best minds, and the best ideas. The essence of Milestone's open platform is to empowerinnovation. We're building more than video management software - we're building an ecosystem that allows partners andcustomers to innovate on top of it," said Sebastian Döllner, VP of Technology Partnerships & Open Platform, Milestone Systems.
Finalist projects
Second place went to Blaize's team for an emergency response tool that uses edge AI to contextualise emergency events and suggest response plans in near real time. Citilog's SmartMap secured third place by integrating incident detection with weather and traffic overlays for improved operational decisions. Other finalists included New Zealand developer Rawinder Singh for RevoFlow, a no-code workflow builder aimed at video analytics, which also won the audience award.
Further shortlisted entries ranged from video analytics workflow tools to AI-driven platforms providing fast video summarisation, advanced search, and semantic metadata extraction.
Industry ecosystem
During the event, industry speakers from NVIDIA, AWS, Dell, and Intel discussed the broader potential of artificial intelligence in video technology, reinforcing a growing focus on open platforms and collaborative development. The hackathon illustrated how domain-specific AI systems may provide avenues for bringing additional value to city infrastructure through automation and managed data privacy.
Data sourcing
Milestone's VLM is trained with Hafnia's data, focusing on compliance and anonymisation to address ethical concerns in video analysis. The approach combines computer vision and natural language processing to convert visual data into structured outputs without identifying individuals.
"We are very impressed with the innovative integrations from all finalists. But Thomas Kreutz and Ask The City really nailed it. The hackathon success is promising for the future use of our platform and data library to train computer vision models on compliantly sourced, curated, extensively annotated, and anonymized real-world data," said Roland Harwood, Community Lead - Hafnia.
