From screenshot to test: multimodal UI validation pipeline

Workflow: 1) Input: Screenshot and DOM snapshot from CI. 2) Vision: Use GPT-4o or Llava for element detection; normalize component names. 3) Heuristics: Map elements to actions (click, type), infer selectors with confidence. 4) Test synth: Generate Playwright spec with step-by-step assertions. 5) Data: Store failures with image diffs and DOM deltas; learn selector fallbacks. 6) Review: Human approves generated tests; feedback loop improves prompts. 7) Run: Integrate into CI; auto-skip flaky steps via retry policy. 8) Metrics: Flake rate, coverage gain, test execution time. Delivers faster regression coverage on complex UIs with less boilerplate.