Video is your largest untapped dataset.

Ask it anything.

Zelf watches video at scale, 1.8 billion clips and counting, and turns it into

answers any team or agent can query.

Zelf watches video at scale,

1.8 billion clips and counting,

and turns it into answers any team

or agent can query.

Read Docs

zelf.ai

watching…

CLIP 01 · PRODUCT720p

00:00 / 00:05CC

FOUND SO FAR0 / 5

watching the clip…

TRUSTED BY TEAMS RUNNING VIDEO AT SCALE

PE FUNDS
PE FUNDS

Video, understood. Not just recognized.

Zelf watches the video itself, not just captions and transcripts.

And it watches all of it:

every video in the set, not a sample.

00:00

f.01

cap twisted open.

00:02

f.02

bottle raised to mouth.

00:03

f.03

drinking.

00:06

f.04

bottle lowered from mouth.

00:08

f.05

cap screwed back on.

00:12

f.06

bottle placed in tote bag.

Ask questions no other system can answer.

Find any moment by describing it.

Track actions and motion.

Detect objects, brands, logos.

Reason over the video, cited to the timestamp.

Interpret against your criteria.

Describe what is happening.

Read on-screen text.

Transcribe speech.

FIND ANY MOMENT BY DESCRIBING IT

00:47 / 00:47

QUERY“the moment they decide to buy”2 hits

00:04“okay, we will take two of those”match

00:08the box changes hands, both lean inmatch

Ask across everything said, shown, or written, in plain English.

01 / 01

Ready for enterprise scale.

1.8B+

Videos processed, about 50M added every month.

100,000

Videos watched in a single batch.

200x

Faster than real-time.

Why not just call a model API?

Raw models break at scale.

Context limits, accuracy decay, no eval loops.

Zelf makes vision AI production-grade.

Batching, accuracy hardening, grounded citations to the exact clip and moment.

Scale without the pipeline.

A raw model takes one video per call. Zelf takes the entire archive in one job and manages the batching, retries and cost behind it.

RAW MODEL · ACCURACY OVER A BATCH

context limitaccuracy decayno eval loop

WITH ZELF · HARDENED PIPELINE

Batchingdone · in flight · queued

Accuracy hardeningheld flat across the batch

Grounded citationsevery claim, clip + second

The price is stated twice, once on the shelf tag clip 0412 · 00:07 and once spoken at the counter clip 0288 · 00:19.

00:0000:0700:1900:30

SCALE · WITHOUT THE PIPELINE

What you would buildRAW MODEL

01Chunk every video

02Queue and rate limit

03Call the model per clip

04Retry the failures

05Merge and store results

06Evaluate accuracy

What you callZELF

POST /jobs

sourcearchive/*

ask"every time someone eats a pizza"

BatchingHANDLED

RetriesHANDLED

Cost controlHANDLED

Eval loopsHANDLED

done

answerscited to clip + second

One engine, three ways in.

Whether you ship a product, write the code, or run a platform, the same vision AI meets you where you already work.

For product teams.

Turn your own video into a queryable asset.

For developers.

Every capability as an API and over MCP, with docs front and centre.

For platforms.

Embed vision intelligence into your product without building the pipeline.

Use cases.

Vision AI for any video. What teams already run on Zelf.

00:00

bottle · 0.96

Brand and product detection

Find a brand or product inside a video, timestamped.

00:08

ball · tracked

Sports and match analysis

Every player, the ball, every event, timestamped.

00:06

subject · track

holds gaze

subject · track

Annotation and motion analysis

Dense motion annotation over time, for frontier training.

@creator_4471

creator

talks to camera

brand-safe · pass

Creator intelligence

What a creator actually does on camera, not the captions.

00:05

the sandal

held · 9s

UX research

Watch someone meet a product for the first time.

01 / 05·

We built the world’s best social video intelligence platform to show the power of the infrastructure.

Explore social video intelligence →

Explore social video

intelligence →

Share of the conversation · 8 months

May · overtakes the category

JanFebMarAprMayJunJulAug

SHARE NOW

34%

8 MONTHS AGO

RANK

01 of 5

How you use it

Ask our agent, connect over MCP, or call any capability in code. Whatever fits how you work.

Build vision AI into your product. Detect brands, annotate footage, vet creators, search the index. One call each.

READ THE DOCS →

Connect your own agent. Zelf becomes a tool it can call, and your agent gets eyes on video.

READ THE DOCS →

Our agent. Ask a question about any video, get a grounded answer cited to the exact clip and second.

<detection-host>

COPY

Detect a brand

Annotate a match

Vet a creator

Search content

cURLPython

curl -X POST https://<detection-host>/v1/detections \
  -H "x-api-key: $ZELF_KEY" \
  -d '{ "rubric": { "targets": [ { "name": "Glow+",
          "identifiers": ["the wordmark on the label",
                          "the amber dropper cap"] } ] },
        "videos": ["https://www.youtube.com/watch?v=..."],
        "mode": "consensus", "options": { "cost_cap_usd": 10 } }'

→ 201 · CREATED · RESPONSE

{
  "job_id": "det_9f2c41ab",
  "status": "queued",
  "mode": "consensus"
}

REAL PATHS · x-api-key AUTH

Claude

ZELF MCP

Across these 10,000 clips, what do people do immediately after opening the package?

TOOL CALL

annotate_actions · 10,000 clips · first 5s after the open

61% smell it · 24% flip to the label · 9% hand it over

A research agent that watches ten thousand unboxings so nobody has to.

WATCHING

61% smell it · 24% flip to the label · 9% hand it over

ONE CONNECTION · EVERY CAPABILITY

PANDORA

↓ EXPORT CSV

Connect Google Drive. Where does our logo appear in these videos, and for how long?

WORKING · GOOGLE DRIVE · 340 VIDEOS

Connected Google Drive340 videos

Watched every frame for the logo47 appearances

Timed each one6m 12s on screen

Placed it on screencup, sign, bag, shirt

Verified against the frames47 of 47

logo · cup · 00:41

REPORT READY47 ROWS

CLIP · SECOND

DURATION

PLACEMENT

Show me the three longest logo appearances.

CLIP · SECOND · WHERE ON SCREEN

Three clips carry most of it: a cup held to camera, a storefront sign, a tote bag on the table.

clip 0288 · 00:41 to 01:03

cup, centre frame · 22s

clip 0117 · 02:10 to 02:26

sign, top right · 16s

clip 0305 · 00:07 to 00:19

bag, bottom left · 12s

Across these 10,000 unboxing clips, what do people do right after opening the package?

FIRST FIVE SECONDS AFTER THE OPEN

smells it · 00:02

Most people smell it first. Reading the label comes second. Only one in ten hands it to someone.

ACTIONSHAREEVIDENCE

Smells it61%6,102 clips, cited

Flips to the label24%2,391 clips, cited

Hands it to someone9%10,000 clips watched

Before we sign these 40 creators, who fits our rules and who has a problem the captions never mention?

CREATOR VETTING · 40 PROFILES · UP TO 150 POSTS EACH

on camera · post 0142

@rae.cooks148 postsPASS

@dan_liftsalcohol on camera · 3 postsFLAG

@mira.beauty150 postsPASS

@theo.techcompetitor unboxing · post 0088REVIEW

31 / 40

PASS

5,812

POSTS WATCHED

Six creators carry a flag the captions never mention.

ANY VIDEO · ANY SOURCE · CITED TO THE SECOND

What will you ask first?

Bring a question about your own video. Leave with the answer.

READ DOCS

Company

Docs

hello@zelflive.com

Privacy

Terms