Everyframe is video software we built and host in Europe. It takes footage you already own and returns finished video at the volume the work needs. Every example below is a real run on real client footage.
One design and one list of entries. Every video came back finished in its own region and voice, with zero manual re-edits.
235 entries / 6 voice variants / 13 regions
1,410
Finished videos
One cell, one rendered video
A title, a cut and an ad are each a TypeScript file. The timing comes from the material. One data file drives every size.
Title storm 8 s 1080 × 1920
Twenty-four letters fly in, each on its own arc. The SDK listened to the song and found the beats. No landing time was typed by hand.
0 landing times typed by hand
title-storm/storm.ts:183
Launch board 10 s Sample data
16:9 · YouTube
1:1 · Instagram
9:16 · Shorts
One JSON file of pre-order numbers feeds all three. Each ad is cut for its platform and kept inside that platform's safe area. Change a number and all three videos follow.
1 data file behind all three, no lint issues
data/board.json:1
369 faces found across 186 frames, all pixelated except the one rider you want on screen. He is held by his track identity, so no face recognition and no boxes drawn by hand. The same run also exports a blacked-out version.
369
Faces detected
186 frames
The engine reads 17 joints per person from the camera already on the floor. It finds the repeating motion on its own, so there is no rule to write per exercise and no hardware to buy, fit or charge. Give it footage that does not repeat and it refuses to count. It will not guess.
17
Tracked joints
No sensor worn
Six more clips the same run counted
A Jetson board on the aircraft does the seeing at 17 W, fully offline. Home gets about 68 bytes per record and never a frame, so the link stays usable on a bad day and the footage stays where the rules say it has to stay.
68
Bytes per record
17 W on the board
34
Frames carried by the filter
out of 250
The lock, frame by frame
Left is the drone feed with the tracker's box drawn on it. Right is the same target magnified. The coloured box is the raw reading for that frame, and the colour says which confidence gate it cleared. The dashed box and the ring around it are the engine's guess and how far off that guess could be. Watch the ring swell when the detector loses the car, then snap tight the moment a reading comes back.
20 frames clear the high gate. 196 are low confidence and kept anyway. 34 have nothing to read, so the filter carries the box. That is all 250 frames, at a mean confidence of 0.38 .
Holding one vehicle is the hard part. A car drops below what the detector can see and the box keeps moving anyway. Confidence gates decide what still counts as the same vehicle, and a filter carries the box forward while there is nothing to read. A solid box is a reading and a dashed box is the engine's guess.
“Perception, not force.” The engine reads a scene and hands it to a person. A human stays in every decision.
Seven passes the engine has run on real footage
Other tools read a JSON template on their own servers every time a video runs. Everyframe compiles a typed description into one portable program. That is compiler work, and it is what every run above rests on.
Models are typed functions. A detector goes from video to detections. If one model's output matches the next one's input they chain. If it does not, the pipeline does not compile.
Mistakes surface when the pipeline is compiled. A batch of 1,400 videos does not fail on the 900th video.
The same compiled program runs in our cloud, inside your own SDK, or on a machine with no network at all.
Nothing on this page is a mockup. Every showcase is a run on footage a client already had, and the numbers beside each one come from that run.