Engineering notes
TangoTube knows who is dancing, to which recording, at which festival, in more than a hundred thousand videos out of four and a half million examined. It runs on one server, with no machine-learning model in the request path, and every fact on it can be traced to the evidence that put it there. Eight short essays explain how, in drawings, for readers who know engineering and nothing about tango.
- videos in the database
- 4,566,186videos in the database
- shown to visitors
- 116,542shown to visitors
- dancers
- 5,039dancers
- with the recording named
- 51,810with the recording named
- server
- 1server
The idea that runs through all eight
Every fact is a clue with a source and a grade. No single clue is ever allowed to make a fact true; two independent clues that agree are. Everything the machine cannot decide is handed to a person rather than guessed. Every human decision becomes a test. And because every rule is deterministic, a mistake, once found, can be traced to the rule that made it, measured, and fixed everywhere at once.
The essays
Read Volume 1 for the whole pipeline, then Volume 8 for the patterns with the tango removed. That is the argument: graded clues, agreement, honest depth, people for what rules cannot decide. The rest is depth. If you would rather see what is coming than how it works, the roadmap is next door.
Counts are read from the live catalog. Figures quoted inside the essays are the measurements that were taken at the time each decision was made, and are written beside the code they justified.