Normal people would probably stick to API for google search. However, sadly the google API was pretty horrible when it came to returning relevant images - in addition to being fairly slow (about a second response time).
Instead, we made a scraper to scrape images from a google search query. Turns out the images were a lot better (more relevant), and we lowered the search to about 300-500ms.
This means that the time from speech -> new slide on a remote screen (via cloud) is pretty close to a second (including overhead).
In addition, we have sensors (see the "This is intense" badge) which streams data to a live HUD in the presentation.