Btw, it doesn't really sound like the problem needs a video as an input to llm. Feels like sending an image is okay. So that makes it less demanding(?)
It's a bit tricky, the fp rate is not ideal, it does wrongly beep from time to time. I haven't really had a serious false negative, but did have some true positives :)
About the hard numbers, it's tough to test it quantitatively, because there's not a lot of data for babies in danger :D and I hope it stays that way
In general, I'm hoping that the open models will get better, there has been a lot of acceleration in video modality recently
Hey! It's fully local, I was trying to build it privacy first. All the things related to kids are very sensitive, I didn't want to send anything to cloud.
But you can still run the inference remotely, changing that should be just a matter of changing the address.
