Research Interests
Neural Dynamical Systems
Treating computation as continuous flow — ODEs & SDEs as a design language for efficient, expressive models.
Continual Learning
Learning as a trajectory with no endpoint — absorbing new tasks over time without overwriting the past.
AI for Scientific Discovery
Turning those same dynamics outward — models that surface structure and accelerate discovery in the sciences.
Robotics on the Edge
Bringing real-time perception and control to robots under tight on-device compute, memory, and power budgets.
About
I'm a Ph.D. Student at UC Berkeley and a Senior Applied Scientist at Amazon Lab126. At Amazon, I work on extreme hardware–software co-design — leading generative-AI capabilities on Amazon's custom silicon to bring on-device GenAI to products like Alexa+. I joined through the acquihire of Nyun AI, the startup I co-founded and led as CTO.
My research explores efficient architectures through the lens of Neural ODEs/SDEs, continual learning, and AI for scientific discovery — building on a track record in model compression, quantization, and efficient foundation models — with papers at CVPR, ICML, ICLR, IJCAI, ACL, and TMLR.
Earlier, I was a founding researcher at Transmute AI Research, a student researcher at Microsoft Research and Google DeepMind, and a research assistant at MBZUAI.
News
- Oct 2026Released Large Flow Language Models, introducing Lafa for efficient parallel language generation.
- Oct 2026Released Subspace Tuning (SST) for mitigating forgetting when fine-tuning language models.
- Fall 2026Started my Ph.D. at UC Berkeley.
- Jul 2026Organized the Efficient LLM Challenge at AdaptFM @ ICML 2026 — 176 teams, with the top entry reaching a 7.7× inference speedup at no quality loss.
- Jul 2026Co-organizing the AdaptFM workshop on Resource-Adaptive Foundation Model Inference at ICML 2026.
- Jun 2026DOT-MoE accepted at ICML 2026, introducing differentiable optimal transport for converting dense models to mixtures of experts.
- May 2026Jacobian-guided Noise Injection accepted at the ICML 2026 AdaptFM Workshop, improving LLM robustness to quantization.
- Feb 2026S2D: Selective Spectral Decay accepted at CVPR 2026, conditioning neural activations for low-bit quantization.
- Apr 2025Nyun AI was acquihired by Amazon Lab126; joined as a Senior Applied Scientist.
Selected Publications
* equal contribution; † equal advising. Full list on Google Scholar.
Experience
Honors
Service
Workshop Organizer — Resource-Adaptive Foundation Model Inference (AdaptFM), ICML 2026; and Resource-Efficient Deep Learning for Computer Vision, ICCV 2023.
Competition Organizer — designed and ran the Efficient LLM Challenge at AdaptFM @ ICML 2026 (176 teams), and authored the results writeup.
Reviewer — NeurIPS, ICML, ICLR, CVPR, ICCV, Pattern Recognition Letters.
Open Source — Lead of Nyuntam, Nyun AI's model-optimization suite.
CV