Accessibility Agents
Redesigning accessibility tools into intuitive, accessible and powerful conversational interfaces.
About the project
Company
Level Access
Timeline
2026
My role
Product designer, AI UX
Challenge
Context
The platform featured fragmented, clunky AI tools. The Reporting Agent relied on rigid and outdated prompt flows, while the Level AI assistant failed to meet modern conversational UX and accessibility standards. The challenge was to overhaul these interactions into seamless, conversational and inclusive experiences that drove business value.
Conversational Architecture
Transitioning rigid forms into natural, vibe-coding style chat interfaces.
Accessibility-First AI
Ensuring streaming LLM outputs met WCAG 2.2 AA standards.
Action-Oriented Flows
Embedding contextual triggers to maintain momentum and close the find-fix-prove loop.
The designs
Due to confidentiality, I can only share a limited view of the project. Here are the core flows and strategies I implemented.
The Reporting Agent Overhaul
Users struggled with conflicting edge-case requests and limited control over the output when generating reports. New AI capabilities lacked accessibility compliance.
What I did
Architected a new vibe-coding style interface featuring rapid prompt refining, robust report versioning and new output control capabilities to smoothly handle complex edge cases.
Impact
Increased generated reports by 23%
Drastically reduced user friction during data extraction.
Reporting agent became fully keyboard navigable
Ask Level AI Chat Upgrade
The Level AI assistant felt robotic and lacked the usability features users expect from modern LLMs. Basic features were missing or hidden, causing severe interaction drop-off.
What I did
Integrated modern AI UX patterns, including contextual feedback triggers, chat history flows, and suggested follow-up actions, transforming it into a natural, accessible conversation.
Impact
Drove a noticeable increase in overall active AI feature usage across the platform.
New action-oriented flows empowered the find-fix-prove loop.
Considerations
Designing for accessible AI Consumption
Conversational UIs often fail accessibility standards. The entire experience was built to keep screen-reader and keyboard users in control; from prompt inputs with explicit keyboard logic to prevent accidental submissions, to aria-live regions to safely manage streaming LLM outputs.
Results
Metrics
Higher engagement
23% increase in report generation and exports.
Increase in daily active users engaging with Level AI features.
Zero accessibility blockers
The redesigned streaming text outputs successfully passed strict accessibility QA for screen reader compatibility.
Next steps
Scale logic
Expand the newly architected reporting logic to support additional enterprise data formats, artifacts, and evolve the chat UI to predict user needs based on dashboard data.
Measure and iterate
Continue benchmark testing with native screen-reader users to validate complex conversational edge cases.
Key takeaways
Novel accessibility for novel AI
Off-the-shelf LLM UI components rarely support WCAG out of the box. We had to architect custom interaction patterns for real-time streaming text, focus management, and dynamic ARIA live regions.
Edge cases define the core experience
When dealing with strict accessibility requirements and conflicting user prompts simultaneously, designing for the edge cases ultimately created a more robust and resilient product for all users.
System state visibility builds user trust
In conversational UX, users need to know exactly what the AI is doing. Designing clear history flows and immediate feedback triggers proved non-negotiable for driving adoption.
Conversation flows reduce cognitive load
Moving from rigid data-entry forms to natural language prompt refining provided a faster, more flexible way for users to handle complex edge cases in data extraction.
