I'm in the process of finding where to put $100,000 of 4090s for training non-llm models.
It looks like "my garage" is cheaper by far even when I include installing an HVAC system and 3 phase power when I look at yearly costs for a spot with 20kW's of power.
You can just train a vision model to detect people, then detect people on a white list. If you're not on the white list you send a head request to the shooter microservice which deals with the rest of the intruders lifecycle.
It looks like "my garage" is cheaper by far even when I include installing an HVAC system and 3 phase power when I look at yearly costs for a spot with 20kW's of power.