Purpose
Demonstrate responsible AI-assisted UX analysis through a small, testable, public product.
← All projects
Production readyPublic repository
Flagship independent project
AI UX Audit Assistant
AI UX Audit Lite is a focused portfolio application that turns one interface screenshot and optional product context into a structured, explainable UX review.
My role
Product definition, UX architecture, interface design, frontend engineering, AI integration, accessibility, validation, testing, documentation, and release preparation.
Current status
Sprint 14 implementation is complete. Sprint 15 production launch preparation is in progress, with final live deployment evidence tracked separately.
The problem
- Manual interface reviews take time and are difficult to structure consistently.
- Generic AI feedback often lacks severity, categories, prioritization, and a dependable response format.
- Screenshot uploads introduce privacy, cost, validation, and failure-handling concerns.
- Automated feedback must remain directional and must not imply formal UX or accessibility certification.
Included in the focused scope
- One screenshot per review
- Optional product context
- Structured summary, directional score, findings, and recommendations
- Severity and category filtering
- Markdown and JSON report export
- Fixture and OpenAI provider modes
Deliberately excluded
- Website crawling or multi-page audits
- User accounts and saved history
- Benchmarking or formal certification
- Persistent screenshot storage
- Commercial monitoring or scoring claims
Product flow
- 1
Upload a supported interface screenshot
- 2
Add optional product and audience context
- 3
Submit the review through the server route
- 4
Validate the provider response against the public schema
- 5
Review summary, score, findings, and recommendations
- 6
Filter, copy, export, or start a new review
System architecture
- Accessible client form
- Next.js server API route
- Provider adapter
- Fixture or OpenAI provider
- Runtime schema validation
- Public response model
- Accessible results interface
Fixture mode returns deterministic schema-valid data without inspecting screenshot pixels. OpenAI mode submits the image and optional context from the server route, then validates the structured response before rendering it.
Key product and engineering decisions
Provider abstraction
Fixture mode supports deterministic, cost-free development and deployment validation, while OpenAI mode can be enabled server-side without changing the results interface.
Runtime validation
Structured provider output is validated before it reaches the interface, preventing malformed AI responses from becoming an uncontrolled UI state.
Directional scoring
The score supports prioritization but is deliberately framed as directional rather than authoritative or certifying.
Privacy-aware boundaries
Provider credentials remain server-side, responses use no-store behaviour, and production guidance prohibits logging screenshots or user context.
Accessibility approach
- Semantic form and results structure
- Keyboard-operable controls and visible focus
- Status and progress communication
- Usable layouts at 200% zoom
- Reduced reliance on colour alone
- Manual and automated quality review documentation
Engineering quality
- TypeScript models and Zod validation
- Vitest coverage for critical logic
- GitHub Actions verification workflow
- File type and 5 MB size validation
- Lightweight request rate limiting
- Production metadata, sitemap, robots, and security headers
What the project demonstrates
- AI output needs deterministic contracts before it can support a dependable product experience.
- A narrow scope improves explainability, privacy control, testing, and delivery quality.
- Fixture mode is valuable for cost control, demonstrations, and safe production verification.
- Accessibility decisions must shape upload, progress, filtering, results, errors, and exports from the start.
Responsible-use boundary
The review is an AI-assisted, directional evaluation. It does not replace user research, expert review, formal accessibility testing, or certification.