restive sparrow
Yeah, finally made it!
Here is my submission for the Nvidia Cosmos Cookbook Hackathon : https://github.com/nvidia-cosmos/cosmos-cookbook/issues/245
- Post-training : Built an app to ease the post-training process. Filmed homemade videos (50) annotated manually in LLaVA dataset format, used also online datasets, and generated synthetic datas (20-40) with Cosmos Transfer. Performed Lora, Qlora and SFT on Nebius H100. Accuracy improved from 35% to 85% to detect behaviors at risk of persons living alone.
- Online app "Angelcare" : An online application was built where the user can connect a RTSP camera, a Telegram Bot and a wearable device (for extra Heart rate monitoring). The models reason and send alerts for behaviors at risk on Telegram for both metrics, and provide Live Summaries + Chat Q&A.
The idea was to build a meaningful project that address two current realities : The AI era, and the aging of the population. 1 in 6 people will be over 60 by 2030 in the US according to WHO/2026, with nearly 3 in 10 adults aged 65+ living alone. The Loom demo only shows a 30 seconds sample of a livestream, but the results show that the post-train model was able to reason over a variety of behaviours at risk.
Thanks to everyone! : @umbral sparrow , @stone rock, @wintry quiver , @sinful temple .
umbral sparrow
sinful temple
stone rock
wintry quiver
clear fiber