Nancy NguyenSenior Product Designer
San Francisco Bay Area · Remote
Currently — Open to senior and staff design roles
Practice: Product · Systems · AI
← All Work
Case Study

Vera AI Platform

Designed the foundational AI interaction layer used by 50K+ daily active users.

RoleStaff Product Designer
Timeline6 months
Team3 Designers, 8 Engineers, 2 PMs, ML Team
50K+Daily Active Users
+38%Trust Score
3xFaster Shipping
4Teams Adopted

Problem

Users interacting with AI-powered tools faced unpredictable, inconsistent experiences. There was no unified interaction model — each feature used its own patterns for input, feedback, and output, leading to confusion, low trust, and poor adoption despite strong underlying model capabilities.

Audit of existing AI touchpoints — inconsistency mapping

Context & Constraints

Vera was an early-stage AI platform serving enterprise knowledge workers. The challenge was designing interaction patterns that felt trustworthy and transparent while the underlying models were still being refined — meaning outputs could be unpredictable, latency varied, and confidence levels shifted.

  • Model response times ranged from 200ms to 12 seconds
  • Enterprise compliance required full auditability of AI outputs
  • Users ranged from AI-native to deeply skeptical
  • Cross-functional team of 12 across 3 time zones

Key Decisions

Confidence-aware UI as a trust mechanism

Rather than presenting all AI outputs with equal visual weight, we introduced a confidence spectrum — outputs styled differently based on model certainty. High-confidence answers appeared definitive; lower-confidence suggestions were presented as options, inviting user judgment. This increased user trust scores by 38%.

Progressive loading with semantic feedback

Instead of a generic spinner, we designed a streaming feedback system that showed users what the AI was doing — "Searching knowledge base...", "Cross-referencing policies...", "Drafting response...". This made variable latency feel intentional and informative.

Composable interaction primitives

Built a system of reusable AI interaction patterns — prompt scaffolds, inline suggestions, conversational threads, structured outputs — that product teams could compose into features without redesigning the interaction model each time.

Solution

A unified AI interaction layer consisting of composable patterns, a confidence-aware output system, and semantic loading states. The platform provided a consistent, trustworthy foundation that product teams could build on — reducing time-to-ship for new AI features from 6 weeks to 2 weeks.

Interaction pattern library — composable primitives
Confidence spectrum — visual system
Streaming feedback states — end-to-end flow

Impact

  • 50K+ daily active users on the platform within 6 months
  • User trust scores increased by 38% after confidence-aware UI
  • New AI feature time-to-ship reduced from 6 weeks to 2 weeks
  • Adopted as the design standard across 4 product teams

Reflection

Designing for AI taught me that trust is the product. When outputs are unpredictable, the interaction layer becomes the primary mechanism for user confidence. The confidence-aware UI was initially controversial internally — some worried it would highlight model weaknesses — but it ended up being the single biggest driver of user trust and adoption. Transparency, done well, builds more confidence than polish alone.

← Back to IndexEnd of study