Scalable Gen-AI Systems for Pro-social Outcome

Platforms define harm universally and respond with suppression. What if users could define harm themselves? We built DIY-MOD, a system with optimized LVLM pipelines that transforms user-defined harmful elements within content in real-time instead of removing the entire post. (CHI 2026)

Current work: Broadening the scope of AI-powered user safety and autonomy.

• Agentic AI for Online Safety: How can autonomous systems protect users across the entire web, beyond specific platforms?
• On-Device Intelligence with WebGPU: Smaller task-specific models (like this) could be augmented in the pipeline to make this pipeline more accessible. Can we tap into that?

User-Steerable Content for Digital Well-being

Short-form algorithms maximize engagement, often leading to compulsive consumption. Can we use Gen-AI to 'cool down' addictive content? We are building DoomsMod, a system that uses on-device VLMs to transform sensory characteristics of videos (e.g., catchy audio, visual intensity) in real-time. It tackles the technical challenge of intervening on encrypted mobile streams to restore user agency.