Snipy, born from an assessment

The update lands quietly: he passed the interview with the end client, and the next gate is a prototype due Monday night ▸ 19:33. The rules are the interesting part: it must be something he’s never worked on before, and what they evaluate isn’t the artifact so much as the process that produced it ▸ 20:01. So the idea he’d been circling since the interview becomes real: a tool that takes these 30-to-60-minute videos and extracts the best parts as vertical shorts, cutting between whoever is talking ▸ 20:23. The name: Snipy, from snip ▸ 21:15.

The hard problem is knowing who’s speaking, and the diary logs each dead end honestly. Attempt one, stereo channels: the recording turns out to be fake stereo, both sides collapse to the same point in space ▸ 21:39. Attempt two, pitch: his voice is lower than Julia’s, and Julia immediately defeats the classifier on camera by dropping hers ▸ 22:26. Julia proposes attempt three, watch the mouths in the frame ▸ 23:03, which he files with the honest caveat that an open mouth isn’t proof of speech.

el problema no es cortar el video; es saber quién de los dos está hablando →

Competitor research keeps the idea alive: Opus, Filmora and friends can track a speaker, but by his reading they can’t track two ▸ 24:06, so the two-person podcast frame is exactly where his tool becomes unique ▸ 24:50. Meanwhile the actual work is Python dependency hell, one package demanding the newest version another refuses, like running Excel 2007 on Windows XP, an uninstall-install loop going nowhere ▸ 25:04. The clock: four days counting Monday, minus the demo video and presentation the deliverable also requires ▸ 26:20…