AI Transparency Tools Fail When Users Can't Evaluate What They See

A plugin for ComfyUI's node-based interface that enables interactive manipulation and inspection of diffusion model internals. Artists can bend and explore different components of generative models to develop intuition about how each part influences output. Supports craft-based explainability through hands-on engagement with model structure for creative applications.

First accessible street view tool combining context-aware multimodal AI, accessible navigation controls, and conversational speech to make Google Street View accessible to blind users. Enables virtual navigation of 220+ billion images across 100+ countries through keyboard/screen reader interfaces. Supports POI investigation, open-world exploration, and remote route planning for blind and low-vision users.

Web-native system for multimodal interaction with 3D surface and point data visualizations, co-designed with blind and low-vision researchers. Provides synchronized auditory, visual, and textual feedback with keyboard and gamepad navigation for exploring 3D continuous scalar fields. Browser-based accessible visualization tool supporting spatial navigation, sonification, and structure discovery for BLV users.

Chrome extension that augments YouTube with in-situ nudges to support descriptive commenting for video accessibility. Provides visual indicators of accessibility gaps, pop-up hints for what to describe, reminders to clarify vague comments, and related captions as references. Leverages Fogg's Behavior Model to engage everyday viewers in creating grounded visual descriptions for blind and low-vision audiences.

Open-source Progressive Visual Analytics tool for exploring aggregated electronic health record sequences at scale. Uses progressive aggregation algorithm to visualize patient sequences as Icicle trees, computing approximate initial results quickly and iteratively improving them. Remains interactive while processing millions of patients—three orders of magnitude more than similar tools—enabling discovery of rare medical conditions.

Augmented reality system for robot instruction through manipulation of spatially-aligned virtual twins of physical objects. Users interact with world-aligned, life-size virtual objects in AR to specify physical goals and spatial parameters. Features real-world lasso selection of multiple objects and snapping to default positions, enabling complex tasks beyond simple pick-and-place operations.

Privacy-preserving ultra-wideband sensing system for continuous, contactless monitoring of ergonomic sitting posture. Leverages commercial UWB devices with comprehensive feature engineering and PoseGBDT model to capture temporal dependencies in posture patterns. Achieves 99.11% accuracy across 19 distinct postures, robust against clothing thickness, additional devices, and furniture configurations for preventive health management.

System integrating large language models and Retrieval-Augmented Generation to identify common misunderstandings in discussion forums and generate actionable teaching insights. Processes student posts to detect misconceptions and create instructional resources. Evaluated with authentic data from three CS courses involving 1355 students and 2878 posts, validated by five instructors for teaching effectiveness.

Live leaderboard platform that ranks LLMs and MLLMs based on human feedback from real AI-powered applications. Employs Bradley-Terry model with Placement Matches for cold-start rating estimation and Proximity Sampling for intelligent model comparisons. Integrates pairwise evaluations into natural user interactions to reflect practical usage scenarios. Publicly accessible at https://www.tbox.cn/about/model-ranking.