Research agenda
AI capability as a full-stack delivery and cognitive acceleration problem.
“Technology is a useful servant but a dangerous master.”
Christian Lous Lange
Research thesis
Whether through AI for Science or physical AI, the goal is to expand the boundaries of human knowledge while preserving human agency. Realizing that goal requires bridging abstract intelligence with reliable, low-latency execution.
01
Four connected layers move intelligence from model capability to dependable real-world outcomes.
Led post-training pipelines and tied model evaluation directly to real workflows, with an emphasis on mitigating hallucination and enabling deterministic, safe execution in mission-critical tasks.
Architected PD-disaggregated inference and explored custom kernel generation, activation pruning, and low-bit quantization for efficient execution across heterogeneous clusters.
Embedded vision-language intelligence into industrial and public-service workflows, combining robust routing with the operating discipline required for commercial deployment at scale.
Research spans stochastic bandits, zeroth-order optimization, graph learning, and noisy-label robustness, grounded in constraints such as gradient noise, limited supervision, and memory bandwidth.
02
Research throughline
Start from a consequential scientific or operational question.
Post-train and align foundation models for the actual workflow.
Co-design inference, kernels, and hardware for reliable low latency.
Measure accuracy, efficiency, safety, and human outcomes in production.
03
Representative work on memory-efficient attention and lossless, verified decoding.
Takeaway: Addresses long-context memory bottlenecks by offloading redundant attention blocks while preserving throughput at massive sequence lengths.
Takeaway: Uses bidirectional tuning to accelerate inference without degrading output quality, balancing speed with model fidelity.
Takeaway: Pairs a lightweight verifier with parallel token generation to reduce end-to-end latency on constrained hardware.