LLM-Based AI Platform
My Role
Design Architect
Duration
7 months
Tools
Figma, Sketch
Overview
Setting User Expectations: Users struggle to interpret the reliability of AI outputs, requiring clear visual differentiation between confident and uncertain responses, along with non-disruptive handling of hallucinations.
Balancing Control vs. Automation: Users want both intelligent defaults and manual steering, creating tension between simplicity and the need for features like prompt tuning, rephrasing, and citation control.
Context & Memory Management: Managing what the model "remembers" can be opaque, with challenges around visualizing past interactions, editing persistent context, and keeping users oriented.
Transparency & Explainability: Users often don’t understand why a result was shown, especially in search or chat; surfacing reasoning or relevance without adding cognitive load is a key hurdle.
Information Density & Layout: LLM responses tend to be verbose, making it difficult to balance completeness with readability—especially when showing sources, follow-ups, or structured answers.
Design for Trust and Transparency: Use progressive disclosure, visual cues, and fallback patterns to clearly communicate AI confidence levels and gracefully handle errors.
Empower Users with Guided Control: Provide structured inputs, follow-up editing, and mode toggles that give users flexibility without introducing cognitive friction.
Simplify Interaction with Contextual Memory: Introduce threaded conversations, editable memory views, and lightweight personalization to make context feel visible and manageable.
Reduce Cognitive Load with Visual Hierarchy: Use chunking, typographic structure, and expandable UI elements to make long or dense responses easier to navigate.
Ensure Cross-Modal UX Consistency: Align patterns, metaphors, and response formats across chat, search, and tools to create a seamless, unified experience.
Test, Learn, and Iterate in Context: Ground design decisions in real-world usage studies, telemetry insights, and explainability testing to ensure solutions meet user needs.
Research
Design
Results
Increased User Trust and Confidence: Transparent cues like citations and confidence indicators led to fewer complaints and higher engagement with AI-generated responses.
Better User Control Without Overload: Guided input tools and smart toggles empowered both novice and advanced users without overwhelming them.
Improved Task Completion and Flow: Context retention and smooth transitions across features helped users complete complex tasks more efficiently.
Reduced Cognitive Load: Chunked outputs, clean layouts, and clear hierarchy made dense AI responses easier to scan and understand.
Greater Consistency Across Modes: Unified patterns and shared UI elements made switching between chat and search feel seamless.
Faster Iteration and More Targeted Improvements: Usage data and feedback loops enabled rapid, focused UX refinements based on real user behavior.
