Designing Trustworthy AI for Legal Decision-Making
Embedded AI into high-stakes legal workflows with scope control, citations, and human oversight to support defensible decisions.
Impact
Increased defensibility by making evidence, citations, and uncertainty visible at the moment of decision-making.
Key contributions
- •Clarified ambiguous questions before analysis to reduce false confidence in AI outputs
- •Designed human-in-control review and override workflows for expert decision-making
- •Unified answers, supporting evidence, and document context within a grid-based review model
- •Made confidence, citations, and explanations visible to support defensibility and accountability
Multi-Platform Design System
Reduced decision churn and rework across 20+ products by establishing shared design foundations and governance teams could adopt incrementally.
Impact
Reduced drift and rework by aligning design and engineering on shared foundations, governance, and adoption practices.
Key contributions
- •Architected a multi-brand token system across 6 verticals from a single source of truth
- •Consolidated 25 acquired products into shared foundations, shipping first unified product in 6 months
- •Built component libraries across web, Word add-in, and Outlook using Token Studio and Style Dictionary
- •Led migration from MUI to Fluent UI without blocking delivery across multiple product teams
Supporting Expert Decision-Making Under Uncertainty
Designed workflows that help experts reason through ambiguity without over-relying on automation.
Impact
Prevented premature trust by separating AI output from judgment and gating automation until expert-reviewed context existed.
Key contributions
- •Structured review experiences to separate AI output from human judgment and final decisions
- •Made uncertainty, exceptions, and incomplete coverage visible rather than hidden behind confidence
- •Designed clear review states and thresholds to ensure expert intervention was intentional and well-timed
- •Prevented premature trust by gating suggestions and automation until sufficient human-reviewed context existed