A minimalist, high-performance design critique tool powered by Gemma 4 on Cerebras.
Bruna UI provides instant, professional-grade UX/UI audits by analyzing screenshots and delivering actionable feedback via a panel of specialized AI agents—all with ultra-low latency.
Upload Screenshot -> WebSocket Transport -> Gemma 4 Analysis -> Categorized Critique
- Zero-Lag Feedback: Built for millisecond-level responses using Cerebras' inference speed.
- High-Signal Audits: Specifically prompted to avoid "AI praise" and provide ruthless, constructive senior-level design critiques.
- Parallel Expert Panel: Instead of a single response, Bruna UI dispatches the analysis to four specialized agents in parallel, streaming a "blind review" from each:
- Accessibility Lead: Focuses on WCAG compliance and inclusive design.
- Visual Architect: Critiques grid alignment, hierarchy, and professional polish.
- UX Psychologist: Analyzes cognitive load and mental model friction.
- Copy Specialist: Mercilessly edits for clarity, brevity, and conversion.
- Stateless Pipeline: Designed for instant "One Frame -> One Response" flow with no database overhead.
- Extreme Minimalism: Only dependencies essential to the visual loop are included.
- Backend: Rails (Minimal mode)
- Frontend: Stimulus + Tailwind
- Transport: ActionCable (WebSockets) for low-latency image transmission.
- AI Model:
gemma-4-31bvia the local Cerebras Ruby SDK. - Data Flow: Base64 encoded image pipeline.
- Ruby installed
- A
CEREBRAS_API_KEY
- Clone the repository.
- Create a
.envfile in the root directory:CEREBRAS_API_KEY=your_key_here
- Install dependencies (Bundler will fetch the exact
cerebrasgem version0.1.0from GitHub):bundle install
bin/devTo ensure your CEREBRAS_API_KEY is configured correctly and the SDK is linked, run the smoke tests:
- Test Basic API Call:
ruby spec/smoke/test_api_call.rb - Test Vision/Image Analysis:
ruby spec/smoke/test_vision_sdk_call.rb
- Live Demo: https://bruna-ui.onrender.com
- Bruna UI announcement on X
- Cerebras Ruby SDK (dependency developed during the same hackathon)
- Google Gemma-4 31B was used to power and develop this project. Model Page on Hugging Face
- Cerebras x Google Gemma 4 Hackaton page
