Lumingyuan Tang

An AI-powered development assistant that 10x your work across UI/UX design, project planning, implementation, and deployment.

Named after Ada Lovelace, the world's first programmer, AdaL is built by SylphAI to enable developers to iterate at the speed of thought. It leverages any model—text, image, reasoning—to accelerate the entire development lifecycle. From UI/UX design to project planning, implementation, deployment, and GTM, AdaL brings 10x productivity with best engineering practices baked in.

An AI agent that autonomously explores literature, explains paper relationships, and frames narratives on complex topics.

Built during the Scoop AI Hackathon (Silicon Valley Bowl), Research Bro tackles the complexity of academic research. It uses a graph-based reasoning engine to connect disparate papers, identifying 'narrative arcs' rather than just keyword matches. The system visualizes these connections in real-time, allowing researchers to 'see' the story emerging from the data.

Built agentic recommendation systems leveraging social graph relationships for personalized commerce experiences.

At TikTok, I worked on next-generation e-commerce recommendation systems that leverage AI agents and social graph analysis. By mining followed-based relationships, we identify optimal relation chains to drive highly personalized and trustworthy product recommendations, moving beyond traditional collaborative filtering approaches.

Bridging LLMs with Knowledge Graphs for deep, responsible, and verifiable reasoning.

Think-on-Graph (ToG) is a framework that enables LLMs to perform multi-hop reasoning on knowledge graphs. By iteratively retrieving and reasoning over graph sub-structures, ToG mitigates hallucinations and provides interpretable reasoning paths, outperforming standard RAG approaches on complex QA datasets.

Enhancing LLM context window utilization through graph-structured memory.

Context Graph introduces a structured memory mechanism that allows LLMs to efficiently manage and retrieve information from massive contexts. By organizing context as a graph, we enable non-linear navigation and retrieval, overcoming the limitations of successful linear attention mechanisms.

A novel framework for adaptive reasoning that allows models to 'ponder' in latent space before generating outputs.

We propose a mechanism where Large Language Models can dynamically allocate compute resources to difficult problems by iterating in a latent mental space. This approach significantly improves performance on complex reasoning benchmarks while maintaining efficiency on simpler tasks.