InfiniteUGC
A production AI video pipeline that turns a plain script into a finished, TikTok-native ad - prompting, AI creator, voice, captions, and edit handled end to end, with no length cap. Runs on AWS infrastructure I built to render, store, and serve video at scale. Used by 50+ brands across 32+ languages.
50+ brands served · 32+ languages · No length cap
the problem
E-commerce brands need a constant stream of authentic UGC-style video ads, but filming with real creators is slow, expensive, and impossible to scale across languages. Generating them naively produces clips that read as synthetic and stop working after a few seconds of watch time.
what i built
Built a multi-stage generation pipeline that takes a script and a chosen UGC style and composes the ad end to end - prompt construction, AI creator selection, voice, caption timing, and edit - then built the AWS infrastructure underneath it to render, store, and serve the finished video at scale. Made it hold quality across 32+ languages rather than only English.
why it's interesting
This was my first system where generation quality was the product, not a feature. Two things decided whether it worked: treating the model as one stage in a pipeline with its own failure modes rather than the whole answer, and building infrastructure that could absorb long, expensive render jobs without falling over. Generative video turns out to be as much an infrastructure problem as a model problem.
built with
next.js, typescript, node.js, postgresql, llm orchestration, generative video, tts, aws