the roast engine.
literally just a brutally honest, highly structured ai diagnostic tool for web interfaces. built for those of us who get a visceral ick from unaligned flexboxes, contrast violations, and overall aesthetic tragedies.
most ai tools sugarcoat their feedback. audacity doesn't care about your feelings. it’s engineered to take a target url, generate a high-res snapshot via a headless browser, and feed it into a vision model. the ai is locked under strict schema constraints to output a structured, actionable, and merciless critique. no yapping, just pure data.
this isn't some basic tutorial project. if you know the standard of my other builds like provia or vidyut, you know this is giving production-ready architecture:
- frictionless pipeline (url-to-image): bypassing cors limits by routing urls through the microlink api. capturing high-res evidence without forcing manual screenshots because that is so 2010.
- strict schema enforcement: the gemini vision model is hard-locked into a rigid json structure using the google gen ai sdk. clean data mapped directly to ui components.
- exponential backoff: engineered to fight through free-tier api traffic jams. the engine silently catches 503 errors and relentlessly retries with compounding delays instead of instantly crashing.
- editorial ui: built with modern css-first tailwind v4 design tokens. stark monochrome layouts, massive typography, very chic.
- frontend: react, vite
- styling: tailwind css v4
- the brain: google gen ai sdk (
gemini-2.5-flash) - capture engine: microlink api
- deployment: vercel
- clone the repo and install dependencies:
npm install