Human Data · company overview

Frontier-grade data for large AI models

Expert-built visual data for vision-language and generative models. Brix finds the Top 1% and puts them to work on your training set.

500+Companies served
120+Countries & regions
4M+Native experts
Backed by HF0 · NVIDIA · Microsoft · Plug and Play
About Brix

Who we are

The AI-native talent infrastructure: global AI recruiting, global teams and Human Data in one system.

500+leading tech companies served

From growth-stage startups to public companies, across AI, engineering, data, design, operations and growth teams.

AIEngineeringDataDesignOperationsGrowth
200k+global community members
50k+Vetted engineers
120+Unicorn AI & tech companies
Remote 101Featured by Lark · #1 remote-work handbook
HF0Plug and Play

Incubated by top programs, building AI-native systems for AI-native companies

Backed by
NVIDIAMicrosoft
Countries & regions

120+ · N. America · Europe · APAC · MENA · LATAM

Serving leading AI & tech companies worldwide
TikTok
Luma AI
Utopai Studios
Fireworks AI
DJI
MongoDB
GMI Cloud
FlowGPT
Tencent
500+
02
Human Data at Brix

AI models trained by Top 1% experts. We build agents to find them.

The scarce input isn't data volume. It's expert judgment.

Expert network
960M+talent profiles indexed
4M+
Native experts
100+
Languages
100+
Domains
500K+Multimodal
150K+Physical AI · Robotics
200K+RL environments & agents
500K+Professional experts
Top 1% by design.Built to reject 99%. These are not crowd annotators.
Why visual-AI teams choose Brix
01

Talent onboarded in days, not months

Recruiting agents and 1,000+ headhunters vet experts from 960M+ profiles. Your team starts within days.

02

Teams and pipelines stood up per industry

Film crews, game studios, creators, engineers: a dedicated team and pipeline for every vertical, tracked live in WorkViz.

03

Deep in visual and vision-language models

Video understanding, captioning, editing pairs, spatial and multi-view data. Expert-reviewed, every delivery.

03
What we supply

Four data product lines

Product lines, flagship datasets and the model capabilities they support.

04
01 · Multimodal understanding

Video clips: livestreams & physics

Clips built around livestream interaction and physical phenomena, with complete events and synced audio.

Live commerce: products

Apparel showcase and narration, keeping the full host-product interaction.

Live commerce: hosts

Dialogue, live performance and audience-facing interaction.

Physics effects & phenomena

Scene effects, object motion and chain reactions.

Scene effects
Chain reaction
05
01 · Multimodal understanding

Video clips: human performance & animals

Song and dance, real emotion, interaction and animal behavior, plus emotionally charged highlight moments of conflict and tension.

Dance
Emotional performance
Interaction
Live performance
Animal behavior
Highlight moment
06
02 · Controllable editing

Video editing pairs

Edit themes include color, objects, product placement, animated text and VFX.

Recolor

Object swap

Product placement

Animated text

VFX

07
02 · Controllable editing

First-person ↔ third-person conversion

The same activity seen from different viewpoints, in solo, duo and group scenes.

Solo
Solo, first-person1st
Solo, third-person3rd
Duo
Duo, first-person1st
Duo, third-person3rd
Group
Group, first-person1st
Group, third-person3rd
08
02 · Controllable editing

Fine-grained local image editing

Remove, replace, add and text edits, checked for target quality and preservation of unedited areas.

Original
Original image

Edit instruction

Add a few dewdrops along the edge of the green leaf at top right; keep everything else unchanged.

Model edit · diff heatmap
Model editEdit diff heatmap

Model output review

Droplet lighting direction and intensity mismatch, weak edge blending, background brightness altered.

Expert fix · diff heatmap
Expert fixFix diff heatmap

Review after expert fix

Target edit, preservation, lighting consistency and blending all pass.

09
02 · Controllable editing

Portrait & lighting editing pairs

Real-shot original, reference and target images expressing changes in viewpoint, pose, tone, depth of field and lighting.

Reference viewpoint

Original A
Original A
Reference B
Reference B
Target C
Target C

Reference pose

Original A
Original A
Reference B
Reference B
Target C
Target C

Contrast

Contrast variantContrast variantContrast variantContrast variant

Lighting · pixel-aligned

Lighting variantLighting variantLighting variant
10
03 · Pro creation & agents

Vlog editing traces

Real creators' footage choices, edit operations and version history, linked to source and final cut.

What a trace records · schematic
Source
Selected
Video
Audio
Subtitle
Versions

Rough cut → revision → final cut, each linked to its parent.

Footage selectionCandidate, used and unused clips, with files, in/out points, rationale and position in the final cut.
Timeline operationsCuts, deletes, reorders, replacements, speed, crops, subtitles and audio, in the order performed.
Version iterationRough cut, revisions and final cut linked, with parent version, timestamps, summaries and reasons.
Structured deliveryJSON for metadata and multi-track info, mapping source to final-cut timing; SRT for subtitles.
11
03 · Pro creation & agents

R2V video editing

Reference-to-video editing: adding a new subject to existing footage.

Original video
Edited video
Subject reference
Rider and motorcycle reference

Key edit instructions

  • Add a rider and motorcycle on the hillside at right.
  • Keep the reference rider's look, gear and bike details.
  • Accelerate, jump, land and ride on.
  • Match camera perspective, terrain contact and scene lighting.
12
03 · Pro creation & agents

End-to-end video generation

One project, one complete package: planning, assets, shot prompts, generation, post and delivery.

Structure A

Video-gen model training package

For video generation, image-to-video, shot generation and quality evaluation.

Character referenceCharacter
Creature referenceCreature
Alien plain scene referenceScene
Generated output

Shots built from character, scene and creature references, with action, camera and sound constraints.

Structure B

Agentic workflow package

For multimodal agents, creative agents and workflow orchestration. Keeps a reproducible process, media, raw data and tooling.

project/ ├─ 1-dataset/ prompts, fields, sample records ├─ 2-media/ final videos, images, previews ├─ 3-raw/ raw requests, responses, graphs ├─ 4-pipeline/ build and processing flow └─ 5-viewer/ browsing and inspection tools
ReproducibleEvery request and response kept, so any shot can be regenerated.
Agent-readyPlanning, asset and shot steps logged as a workflow an agent can learn.
InspectableA bundled viewer for browsing shots, prompts and outputs side by side.
13
04 · Gaming & simulation

Blender 3D grey-box pairs

Reference video on top, the matching grey-box (blocking / layout) animation rebuilt in Blender below. Editable 3D motion and camera from 2D footage.

Single · basic actions
Single · athletic moves
Multi-person · complex
What we build

Character motion and camera moves recreated in Blender, paired frame-for-frame with the reference.

Coverage

Basic single-person actions, continuous athletic motion, complex multi-person interaction.

Quality bar

Timing, key poses and framing match the reference; motion stays natural and continuous.

Delivered files

.blend · .fbx
animation.mp4
side-by-side.mp4 · readme

14
04 · Gaming & simulation

Gaming & simulation data

Gameplay video, control input and metadata in sync, linking observation, action and environment feedback.

Gameplay video
Controller input
Capture types
Free roamRevisitRoam + combatCamera-rotation roamWith commentaryNo commentary

Each session: screen capture with synced audio, plus button and stick events, timestamped.

Sample recordable titles · single-player
Ghost of TsushimaGTA VElden Ring
Red Dead Redemption 2Cyberpunk 2077The Witcher 3
Hogwarts LegacyFallout 4Horizon Forbidden West
Horizon Zero Dawn Rem.Avatar: Frontiers of PandoraThe Elder Scrolls IV
Grounded / Grounded 2The ForestSons of the Forest
PalworldIcarusAstroneer
Crimson DesertFinal Fantasy XVDragon Age: Inquisition
Ghostwire: TokyoSky: Children of the LightSable · Tchia · TerraTech

Multiplayer: It Takes Two, Split Fiction. Customizable and growing.

Supporting data
raw_meta_data/ └─ 2026-04-27_12-00-58/ └─ input_capture.json raw_videos/ └─ 2026-04-27 12-00-58.mkv
  • Key mapping and timestamped events
  • Input linked to video frames and task logs
  • Task metadata per session
15
04 · Gaming & simulation

Multi-camera game data

Coordinated multi-camera movement in one 3D scene, capturing synced multi-view video and camera poses.

4-camera synced sample

Coordinated cameras

In high-quality 3D games, multiple cameras follow set paths, moving through and exploring the same scene together.

Strict sync & alignment

Multiple views recorded at the same instant; frames and camera poses aligned to the millisecond.

session/ ├─ action_camera_1…4.json camera × 4 └─ systeminfo_1…4.json system × 4
16
Working together

Three ways to work with us

Engagement
You provide
We deliver
01

Data solution design

Needs still to be scoped: define the data plan first.

Model goals, samples, existing data and constraints

Data schema, production standards and acceptance plan

02

Off-the-shelf datasets

Existing products fit: evaluate and license.

Use case, specs, volume and license scope

Matched datasets, field docs and license files

03

Custom data production

Task and specs are clear: run a dedicated production.

Task definition, samples, volume and acceptance rules

Pilot samples, batch data and QA reports

How we kick off
Align on needsShare target tasks and data specs
Validate samplesReview samples, calibrate standards
Pilot acceptanceAgree pilot scope and acceptance
Batch deliveryDeliver and accept in agreed batches
17

Let's build your next dataset.

Tell us about your multimodal training and evaluation data needs.

MultimodalEditing pairsAgent workflowsGaming & 3D
katherine@joinbrix.com joinbrix.com
18