Designing for AI You Can Trust
Backlit AI is building infrastructure for knowledge and obligation management -- a category where the cost of missing something is high and the information is inherently complex. As the founding product and UX lead, I own the design of the entire platform experience from scratch.
The Core Challenge
AI-powered tools fail users not because they give wrong answers, but because users can't tell when to trust them. The central design problem at Backlit was building an interaction model that made AI outputs legible, verifiable, and actionable -- without overwhelming the user or hiding the system's reasoning.
What I Designed
I developed a structured state-transition model (Open / Waiting / Resolved) that gives users a clear mental model of where the system is in its reasoning process at any given moment. Inference feedback loops let users see not just what the AI concluded, but what it considered. Explainability patterns surface confidence signals and source attribution in context, so users can make informed decisions without leaving the workflow.
Design System
Built a full component architecture from the ground up, including tokens, interaction conventions, and documentation designed for a small, fast-moving engineering team. The system was built to scale with the product, not just to serve current screens.
What It Required
This work demanded deep collaboration with engineering on API behavior and data model constraints, a high tolerance for ambiguity in an early-stage product, and a strong point of view on how to communicate uncertainty clearly -- a skill that transfers directly to any platform where system state matters.
