Petabytes of Multimodal Data

We provide millions of hours of proprietary computer use, talking head, audio, video game, annotation, and expert data to power the next generation of multimodal models.

01 / What we sell

Three ways to get video, depending on how much you need it changed.

Most teams start in the store and come to us later, when the thing they need does not exist yet and somebody has to go make it.

RecipeBook

A self-serve store. Type what you are looking for, vote clips up or down, and a small model re-ranks the rest to match. Order the batch by the hour at $3. Filter by quality, length, frame rate, aspect, clip type, transcript, and caption.

recipebook.shofo.ai

Custom datasets

Built to order. You tell us what the model has to learn; we find the footage, clean it, cut it into clips, and label it. We have built cooking videos with hand-and-object detail, egocentric and driving footage, robotics, talking heads, and gameplay paired with the inputs that produced it.

Tell us what you need

Open datasets

Free on Hugging Face, so you can see how we cut and label before you spend anything. The first one is 10,000 English talking-head clips, about 186 hours.

huggingface.co/Shofo
02 / How RecipeBook works

Four steps, and the search learns as you go.

The point of the probe is that you do not have to write a perfect query. You show the search a few clips you like and it does the rest of the sorting.

  1. 01

    Search

    Type what you want in plain words.

  2. 02

    Vote

    Thumbs up or down on the clips you see.

  3. 03

    Train a probe

    A small model learns your taste and re-ranks 25M clips.

  4. 04

    Order

    Buy the batch by the hour. Pay by card.

the probe sends you back to a better search

03 / The library, counted

What is actually in there, as of September 2026.

These are our own counts, taken from the index that RecipeBook searches.

25M+clips in the library
255,000+hours of video
3,899source datasets
271TBof source video held
3.2Mclips with a transcript
23.7Mclips with a written caption
36video embedding models we benchmarked, across 73 benchmarks
$3the price of one hour of video
Backed by Y Combinator · W26
04 / One clip, in full

Every clip arrives with this much written down.

This is a single record, copied out of a cooking set we built earlier this year. You get the same fields for all 25 million.

clip_8f3c21ae4d
record 1 of 96 in batch b_44127
Duration
00:00:12.480
Frame rate
29.97 fps
Resolution
1920 × 1080 · 16:9
Clip type
real
NSFW flag
false
Source dataset
ds_3117 — “Home cooking, handheld, 2025”
License
CC BY 4.0 · attribution carried on the record
Caption

A pair of hands dices a yellow onion on a wooden board, sweeps the pieces up with the flat of the knife, and drops them into a hot pan at the right of frame.

Transcript

“…get the pan properly hot before anything goes in, otherwise the onion just steams and you lose the colour.”

05 / Why we exist

Buying video should feel like buying anything else.

Getting video for a model is still slow. You email a vendor, you wait, somebody sends a quote, you sign something, and a few weeks later a drive shows up full of clips that are almost but not quite what you asked for. We had that job before we started this company, and we did not enjoy it.

So we put the library behind a search box and a card reader. If what you need is already in there, you should have it in an afternoon. If it is not, that is the work we like best, and you can hand it to us instead.

Braiden Dishman, CEO · Alexzendor Misra, CTO · Andre Braga, Chief Scientist
06 / Start

Tell us what you are training.

Send a few lines about the model and the footage it needs. We will read it and set up a call. If you would rather just look around first, the store is open.

Record for us — earn.shofo.ai