Inspiration
A form can look usable while keyboard navigation skips the control needed to finish. For someone relying on a keyboard, that can turn a simple booking or registration into a blocked task.
AccessProof explores a practical question: can an AI agent attempt a specific keyboard task and give developers evidence they can inspect, reproduce, and retest?
Its intended users are campus web teams and small organizations maintaining essential forms.
What it does
AccessProof attempts a booking task using only keyboard actions in a real browser. The tester supplies a goal and the confirmation details expected when the task succeeds.
The application records keyboard actions, focus changes, field values, screenshots, and an action timeline. Developers can download a JSON report, a Playwright browser trace, and a standalone reproducer of the recorded actions.
The browser independently checks the visible confirmation. The model cannot make a task pass simply by declaring success. An unfinished attempt without evidence of a specific barrier remains inconclusive, while execution errors are reported separately.
How it was built
Live mode uses nvidia/Nemotron-3.5-Super-VL through Nebius Token Factory. The model receives the task, visible control information, current focus, field values, and recent action history. It chooses one validated keyboard action at a time.
Playwright executes that action and returns the updated browser state. Screenshots are captured as evidence for developers to inspect.
The application uses Python, FastAPI, Playwright, SQLite, and an HTML/CSS/JavaScript dashboard. Render hosts the application and browser worker; NVIDIA model inference runs through Nebius Token Factory.
Token Factory’s OpenAI-compatible interface allowed integration without provisioning a local GPU or downloading model weights.
Challenges
Native dropdowns required careful handling: selecting the observed option label and pressing Tab to commit the choice. Recording selection values before and after each action helped diagnose this behavior.
Another challenge was separating model execution failures from website barriers. Strict action validation and safe response diagnostics help keep these outcomes distinct.
Demonstrated results
A recorded hosted live sequence used the same fictional booking details
Log in or sign up for Devpost to join the conversation.