There's an object detection library (YOLO[1]) freely available that achieves similar bounding boxes/matches to ambient's homepage demo; Joseph Redmon also gave a TedX[2] about it that describes some of the technical details.
Really great idea, thanks for sharing! I do wonder, is this really annotated feeds?
For me, this is machine-curated feeds rather than a form of annotation given no additional information (besides meta-data scrapped and categorised) is displayed. Not quite as catchy, but annotations make me (at least) think of something else.
Note: am passionated about designing to support annotations on media directly.
[1]: https://pjreddie.com/darknet/yolo/
[2]: https://www.ted.com/talks/joseph_redmon_how_a_computer_learn...