The data engine for computer-use AI
Sarsa captures how people actually work inside software, generates meaningfully different ways of finishing the same tasks, and verifies every trajectory by executing it.
Book a demoThe data engine for computer-use AI.
One human demonstration, many verified ways to finish the same task.
The data enginefor computer-use AI.
Real work, recorded faithfully.
Clicks, keystrokes, and application state. The full texture of how a person finishes a task inside real software.
One demonstration becomes many.
Our research team's generation methods produce new trajectories that finish the same task in meaningfully different ways.
Proven by execution, not inspection.
Every generated trajectory runs inside the real application and is checked against the task objective. Failures are discarded.
5 verified · 1 discarded, from one demonstration
The problem
Human demonstrations don't scale.
Computer-use agents learn from recordings of people working. Those recordings are slow to produce, they demand expert time, and each one teaches a single way of finishing a task. As agents get more capable, the gap between the data they need and the data humans can record keeps widening.
Sarsa's premise: if one trusted human trajectory can produce many verified trajectories, each a different valid way to finish the same task, then the cost of diverse training data falls toward the cost of verification.
How it works
Capture
High-fidelity recordings of real computer work: what the user sees, the actions they take, and how the application responds.
Generate
Trajectory-generation methods from our research team turn each demonstration into new, valid ways of completing the same task. Different strategies, different states, same intent.
Verify
Generated trajectories are executed inside the real software and checked against the task objective. What doesn't hold up is discarded.
For labs
Three ways to work with Sarsa.
Licensed datasets
Verified human and generated trajectories for training, fine-tuning, and evaluating computer-use models.
Custom generation
Specify the applications, workflows, and task distributions you need. We produce a verified corpus to that spec.
Infrastructure On the roadmap
Our capture, generation, and verification stack pointed at your own environments, producing proprietary data continuously.
Why now
Every major lab is training agents that operate software. All of them need more interaction data than humans can record.
We started with spreadsheets and everyday computer work, environments where correctness can be checked exactly. The same capture, generation, and verification layer generalizes to the rest of the desktop. Sarsa's aim is to be the data foundation that general computer-use agents are trained on.