He had the audience and a waitlist before there was an app. What he needed was something real enough to raise on.
The situation
A creator with a substantial following already had people waiting for the product, and investor meetings were coming up.
The problem was that the app didn't exist yet.
That meant there wasn't much room for a prototype that only looked convincing. Investors needed to see the actual product working, but there also wasn't time for a long development cycle. We had roughly one to two months to get something real in front of them.
The media pipeline
A short-form video app is more complicated than it looks.
Feeds, profiles and follows are fairly standard. The harder part is everything that happens between someone pressing record and another person watching the finished video.
Editing happened on the phone using FFmpeg. Trimming, audio, watermarks, stickers and text overlays all had to happen on-device.
I hadn't worked with FFmpeg before this project, so a big part of the work was learning how to use it properly on a phone. It wasn't enough for a command to work. It had to work within the phone's memory and battery limits, and it couldn't leave users staring at a loading screen for ages.
Video delivery also needed to be reliable. Uploads went through a CDN with caching so that people weren't constantly waiting for videos to buffer.
Trust and safety
Because users could upload their own videos, moderation had to be part of the app from the beginning.
NSFW detection happened on the device before a video was uploaded. That meant content that failed the check never had to leave the phone, which reduced the amount of server-side processing and meant we weren't storing content that we already knew we wanted to reject.
The detection model was a TensorFlow Lite model bundled directly with the app.
Audio created another problem. Users could add music to their videos, so we couldn't just trust the filename or whatever information came with the upload. Audio was checked against existing music to identify potential copyright matches.
Messaging sat on top of the rest of the system as well.
Text on video
One of the problems that took much longer than expected was positioning text on videos.
While editing, a user could drag a caption anywhere on the screen. When somebody else watched the video, that caption had to appear in the same place.
The problem is that the editor and the final video don't necessarily have the same dimensions.
A user might edit on one preview size, while the actual video could be a different resolution and then be played on another phone with a completely different screen size.
So the position had to be converted between those coordinate systems without making the text move around between the editor and playback.
Getting that conversion right, across different video sizes and devices, took far longer than I expected.
Where it is now
The MVP shipped in time for the investor pitch, and the funding round closed.
The product eventually moved in a different direction, so the app itself is no longer active. That's fairly normal for an MVP. It did what it needed to do at the time.
We had roughly one to two months, with the deadline being driven by investor meetings rather than a normal product roadmap.
A lot of the engineering was also invisible from the outside. A video feed can look simple whether or not the video behind it encoded quickly, played smoothly, was delivered efficiently, and had been checked before anyone else saw it.
That was most of the work on this project.