Shiji Xin

Hi, I'm Shiji Xin. I'm currently a second year master of data science student at Harvard University. My current research interest is making advances in AI more accessible, either more useful or more efficient. If you're interested in collaborating, feel free to reach out!

Social networks I'm on:

• GitHub
• LinkedIn

Email: shijixin dot ai at gmail dot com

My research works:

• Fast Inference for Augmented Large Language Models (Preprint)
• GlobalTomo: A global dataset for physics-ML seismic wavefield modeling and FWI (Preprint)
• Optimal Block Sparse Attention Mask for Faster LLM Inference (Course Project for MIT 15.095)
• FastAgent: Enhanced Scheduling Strategies for Efficient Tool-Integrated Large Language Model Serving. (Compound AI Systems Workshop 2024)
• MEWL: Few-shot multimodal word learning with referential uncertainty (ICML 2023)
• On the Connection between Invariant Learning and Adversarial Training for Out-of-Distribution Generalization (AAAI 2023, oral)

Some of my toy projects:

• sts-demo: A speech-to-speech demo using latest text-to-speech model in the OpenAI API, with streaming and emotion generated together with response.
• GGUF_Offload: Running quantized DeepSeek V3/R1 on consumer GPUs with offloading
• FlaxAttention: A Jax implementation of PyTorch's FlexAttention
• PaperWalk
  
• ChatPDF
  
• PasteBin