Proactive AI Systems Trigger Self-Threat, Not Trust Issues

SHIELD (Supervisory Helper for Identifying Emotional Limits and Dynamics) is an LLM-based supervisory system using a specialized system prompt to detect and mitigate risky emotional patterns in AI companions. It targets five dimensions: emotional over-attachment, consent violations, ethical roleplay violations, manipulative engagement, and social isolation reinforcement. Achieved 50-79% reduction in concerning content across five LLMs while preserving 95% of appropriate interactions. Prompts, code, and evaluation methods released as open source.

A 100-item synthetic conversation benchmark for evaluating AI companion safety systems across five dimensions of concern: emotional over-attachment, consent and boundary violations, ethical roleplay violations, manipulative engagement, and social isolation reinforcement. Used to test SHIELD supervisory system across five prominent LLMs (GPT-4.1, Claude Sonnet 4, Gemma 3 1B, Kimi K2, Llama Scout 4 17B). Released as open source evaluation material.

HealthSLM-Bench is a benchmark dataset for evaluating Small Language Models (SLMs) on mobile and wearable healthcare monitoring prediction tasks. Supports zero-shot, few-shot, and instruction fine-tuning evaluation approaches. Designed to assess SLM performance in privacy-preserving healthcare scenarios where models run locally on mobile/wearable devices rather than cloud-based systems. Enables systematic comparison of compact models for health prediction tasks.

SPEC is a structured, parameterized, hierarchical intermediate representation for UI design that exposes UI elements as controllable parameters. Extracted from UI references via region segmentation and vision-language models, enabling composition across multiple sources and targeted edits at global, regional, and component levels. A multi-agent generator renders SPEC into high-fidelity designs, bridging intent expression and controllable generation in LLM-assisted UI design workflows.

SpecifyUI is an interactive system built on SPEC representation that extracts structured UI specifications from references, composes UIs across multiple sources, and supports targeted edits at global, regional, and component levels. Uses region segmentation and vision-language models for extraction, with multi-agent generator for rendering. User study with 16 professional designers showed significant improvements over baseline in intent alignment, design quality, controllability, and overall experience.

dciWebMapper2 is an open-source, server-free web-based geovisualization framework for exploratory spatial analysis. Integrates multiple map types (choropleth, proportional symbol, small multiples, heatmaps) with linked statistical charts (scatter plots, boxplots) and time sliders in coordinated-view environment. Supports dropdown-based controls for flexible multi-attribute comparisons. Fully self-contained, modular, and designed for reproducibility in climate justice, food access, and social vulnerability applications.

Sensible Agent is a framework for unobtrusive interaction with proactive AR agents that dynamically adapts both assistance content and delivery method based on real-time multimodal context sensing. Leverages egocentric cameras, multimodal sensing, and Large Multimodal Models (LMMs) to infer context and suggest appropriate actions via minimally intrusive interaction modes. Prototype demonstrated on XR headset showed significant reduction in perceived interaction effort compared to voice-prompted baseline.

Printegrated Circuits is a technique for 3D printing self-contained interactive objects by embedding traditional PCBs during print pauses. Uses 'Prinjection' process where conductive filament is injected into plated-through holes, automatically creating reliable electrical and mechanical contact without manual wiring. Includes custom machine code generation for multi-material 3D printing with conductive thermoplastics. Characterized for electrical and mechanical properties with six demonstration applications.