1st place
NVIDIA DGX Spark™
Friday, October 2 · San Francisco, CA
Build agents that unlock the value of video. Explore video understanding, semantic search, and real-time detection, then show a working prototype at the end of the day.
Opens 10:00 AM PT · Deadline 4:30 PM PT
One project per team · Up to 4 people

Build a video agent that searches footage, answers questions, spots events, or takes action. One idea, working by the end of the day.
Pick a specific use case: “flag someone missing a hard hat,” rather than “watch for safety issues.” Adapt your idea to the footage available on build day.
Search depends on what the model’s prompt captures in each video description. You can ingest video with a different prompt on build day.
A working website and project screenshot are welcome. Build your project during the event.
Pacific Time · Event details on Luma
Join the Cosmos Community to access the VAST build environment, create your Cursor account with your registration email for event credits, and create a Weights & Biases account for inference. Bring a laptop and charger. Teams can have up to four people.
Build-environment access is limited to the first 100 attendees to arrive. Additional builders are welcome to join and connect while space is available!
1st place
NVIDIA DGX Spark™
2nd place
Hugging Face Microduck
Two teams win overall hackathon prizes. Credits and gift cards are split among all team members.
VAST AI OS powers video ingestion and keeps vectors, metadata, and video data together.
Build your video agent in Cursor. Sign up with the email you used to register so your event credits can be added.
CoreWeave provides the GPUs. Weights & Biases provides serverless inference, experiment tracking, and observability. Create an account before the event.
NVIDIA's video search stack uses VSS Blueprint and Cosmos to turn video segments into searchable descriptions and embeddings.