Projects
Selected systems and learning infrastructure. Each project is presented as a problem, an implementation, and an outcome.
Concept2Fable
2026Traceable mechanism modeling and structural analogy evaluation
A bilingual K-12 knowledge-graph and LLM pipeline for mechanism-constrained statements, educational narratives, and cross-domain fables. It curated 6,541 mechanism graphs into 3,183 bilingual three-text units (19,098 texts), then evaluated 11 models on three structural-analogy tasks.
Minimal sufficient context for coding agents
A research and engineering track for adaptive compression, long-term memory, and conflict-aware updates. The implementation direction organizes development history, code, and documentation into traceable context for bug fixes and feature work.
ContextLens
2025–2026AI-powered mixed-reality learning system
A Meta Quest 3 prototype that unifies course video, e-books, voice, notes, and physical-scene context. Its Context Tree makes multimodal inputs traceable for grounded LLM answers; the project received a Challenge Cup Silver Award at Xiamen University.
Handwriting-led historical exploration
A tablet-first PWA that preserves handwriting and connects transcription, historical figures, works, and source verification. Marginal notes, evidence cards, and stamps distinguish sourced facts from model associations.
Ω Desktop Pet
2026AI companionship and desktop interaction prototype
An Electron, React, and PixiJS desktop pet prototype with a transparent floating window, conversation, emotional state, memory summaries, and local persistence. The automatic-diary development branch has been merged upstream.
LLM.learn
2026Course notes and learning knowledge base
A course-centered knowledge base connecting official resources, lecture notes, explanations, bilingual transcripts, and PDFs. Includes CS336 2026, CS149 Fall 2025, CMU AI Agents, and agent and infrastructure topics. Public content is reviewed. Source repository.