I'm an AI researcher and engineer working on multimodal generation, full-duplex speech interaction, and efficient inference. With a background in computational mathematics, I have over a decade of experience taking AI systems from model design and training to optimized runtimes and production applications.

My work spans speech, image, and video generation. I've led the development of end-to-end Speech LLMs with streaming dialogue, real-time interruption, multi-turn context, and zero-shot voice cloning. For full-duplex interaction, I worked on jointly modeling listening and speaking over time, allowing the model to learn when to speak, pause, and take a turn. My speech work also covers recognition, assessment, and synthesis, including large-scale TTS systems and implementations of WaveNet, Tacotron, WaveRNN, Parallel Tacotron2, VALL-E and Moshi.

In image and video generation, I've worked on virtual try-on and conditional image editing with Flux Kontext. My video work includes WanVideo and HunyuanVideo fine-tuning, DiT and flow matching, and conditioning methods for clothing consistency and controllable motion. These projects involved data preparation, model training, and inference optimization for practical content-generation workflows.

I work across the training and inference stack: PyTorch, Megatron-Core, and FSDP for distributed training; CUDA, cuBLAS, AVX, vLLM, and quantization for efficient execution. I also build audio capabilities for AI agents, including understanding, generation, and context integration. I care about the details that make models useful in practice: output quality, memory use, latency, and reliable deployment.

I share research notes and engineering lessons on X and my blog. I'm happy to connect with people working on multimodal generation, real-time voice interaction, and efficient AI systems.