CUA-S1 Form Automation

Overview & Result

CUA-S1-FORMS demonstrates how a compact, non-generative System 1 architecture solves tedious browser interactions. Instead of sending multi-megabyte screenshots to multimodal LLMs, it parses the DOM and maps user records into input targets in a single 50ms pass.

How Jev fits in the loop

1. Inspect the DOM for interactable form controls, labels, and placeholders.
2. Match candidate fields against profile/resume key-value pairs.
3. Generate typed DOM fill operations without hallucinating unrequested fields.
4. Allow the user to review the prefilled form before final submission.

How to reproduce

1. Clone the trycua/cua repository and install dependencies.
2. Provide input JSON containing applicant or profile information.
3. Run the form-fill evaluator against a target HTML form to verify automated field matching and fill accuracy.

Why this build matters

Directly solves the slow, expensive problem of screenshot-based computer use for structured browser automation.

Reported performance

Latency: ~50 ms per form fill pass

Limitations

• Optimized for standard HTML form inputs and select elements; complex canvas-rendered inputs require vision fallbacks.
• Edge cases with conflicting labels or nested iframes require deterministic validation before submission.
• Irreversible submit actions must require human confirmation in production.

Patterns