I am Jeng-Yue (Buffett) Liu 劉正悦, a Master’s student in Artificial Intelligence and Innovation at the School of Computer Science, Carnegie Mellon University. My work sits at the intersection of efficient machine learning systems and generative models, building models that are not only capable, but fast, controllable, and reliable enough to run in the real world.

At the Music and Audio Computing Lab, Academia Sinica, advised by Prof. Yi-Hsuan Yang and Prof. Li Su, I proposed SynthCloner (ICASSP 2026), a factorized codec disentangling timbre, content, and ADSR envelope for controllable synthesizer preset conversion, cutting multi-scale STFT loss by 47.3% over state-of-the-art baselines. I then carried that line of work into production at Neutone Inc., hardening their real-time tone-morphing plugin against out-of-distribution audio.

At CMU, my project work centers on making inference fast and dependable. I built grammar-constrained decoding for diffusion language models, along with CUDA sparse-attention kernels that reach a 22–50× speedup on NVIDIA B200. I also shipped Hypoll, an AI-powered social platform for real-time voice streaming, to iOS. Before CMU, I earned dual Bachelor’s degrees in Information Management (B.B.A.) and Geography (B.S.) from National Taiwan University, graduating summa cum laude in the top 1% of the school.

My interests center on:

• Efficient LLM inference & serving systems: constrained decoding, sparse attention, GPU kernels
• Diffusion and discrete-diffusion language models
• Controllable music and audio generation: timbre / content disentanglement