Inference / Analytics•Research Thread
Structured data from the web
We turn messy crawls into normalized, queryable signal for inference and analytics studies.
Research
Three lines of work to make agents fluent in reality.
We watch how humans really work, clean the signal, and hand it directly to inference, post-training, and RL teams.
Focus areas
POST TRAINING•Research Thread
The largest dataset of human computer-use trajectories
Opt-in desktop captures show how real operators plan, navigate, and correct — perfect post-training fuel.
POST TRAINING•Research Thread
Computer Use RL Environments at Massive Scale
Full-stack virtual desktops let agents practice computer use end-to-end with measurable rewards.
How we run studies
- Collect: Consent-first instrumentation captures human-computer signal with provenance.
- Understand: Lightweight analytics and review loops surface the reasoning gaps that matter.
- Deploy: Structured drops and RL environments plug straight into evaluation or post-training.