HUMXN
An AI and robotics data company. HUMXN captures and sources expert human data, and builds reinforcement learning environments, for AI labs and robotics companies.
The problem
Models learn from human data, and the way AI gets it is broken: builders inherit legal risk, creators are taken from without being asked, and the experts whose judgment models need most are rarely in the loop.
Robotics has it harder still. There is no internet-scale archive of people doing physical work from their own point of view. That data has to be captured.
What I built
HUMXN sources datasets to order. Contributors capture first-person video with head- or chest-mounted cameras, screen recordings, voice, expert demonstrations and robot teleoperation, and experts such as physicians, lawyers and engineers annotate and evaluate it.
Every file carries its provenance: hashed, listed in a signed manifest and timestamped, so a buyer knows where the data came from and that it was licensed. It ships in the formats training pipelines already read.
What it does
Data capture
First-person video, screen recording, voice, sensor capture and robot teleoperation.
Expert annotation
Physicians, lawyers, engineers and linguists labeling and evaluating data in their field.
Every training stage
Pre-training, SFT, RLHF and preference data, evaluation sets and multimodal data.
RL environments
Reinforcement learning environments for AI labs and robotics companies.
Provenance
SHA-256 and perceptual hashes, Ed25519-signed manifests and C2PA where the format allows.
Ready to train
Delivered as JSON Lines, Parquet, WebDataset, COCO or Croissant, by API.
Under the hood
- Egocentric capture rigs
- Robot teleoperation
- C2PA
- Ed25519 signing
- Parquet
- WebDataset
- Stripe payouts
Want a system like this working for your business?
I design, implement, and optimize agentic systems that generate revenue, automate operations, and scale businesses. Tell me what you're working on.









