Notes from the RoboEdit team
Releases, engineering write-ups and research from working with 180,000 creators.
Why we rebuilt caption alignment around phonemes
Word-level timing looked fine until we tested Japanese and Finnish. Here is how phoneme-level alignment took accuracy from 94.1% to 98.6%.
The 90-minute podcast workflow used by 400 shows
A repeatable pipeline: ingest, silence removal, speaker switching, clip ranking and eight vertical exports before lunch.
GPU scheduling: how we keep 4K renders under three minutes
Queue fairness, preemption and warm model pools across 1,900 H100s — plus the mistakes that cost us a bad week in April.
Localization economics for creator businesses
We analyzed 2,300 channels that added dubbed tracks. Median watch-time lift was 34% within 60 days.
Designing an editor that trusts the user, not the model
Every AI action in RoboEdit is a reversible layer. Here is the design rule set behind that decision.
RoboEdit 4.0: avatars, dubbing and the new render queue
The largest release since launch, including lip-synced dubbing and a rewritten queue UI.
Get the monthly RoboEdit dispatch
Product releases and workflow teardowns. One email a month.