A catalogue of the research systems I have built or co-built during my PhD, alongside earlier undergraduate work and creative side projects.
A speculative-decoding framework that accelerates reinforcement-learning rollout generation for large language models. A length-aware speculation policy prioritizes aggressive decoding on long-tail trajectories, reducing rollout makespan by up to 50% on agentic reasoning workloads with no change in accuracy.
Low-latency workspace versioning for computer-use agents, built on Linux kernel modifications, CRIU extensions, copy-on-write memory sharing, and filesystem versioning. Enables fast branching of live GUI workspaces with isolated process, memory, filesystem, and GUI state. Evaluated across 600+ agent tasks, achieving up to 1.9× lower end-to-end task latency over KVM baselines.
A Rust-based global scheduler with efficient Python/C++ local schedulers and custom CUDA kernels, treating tensor parallelism as a control surface to dynamically meet strict per-tenant SLOs across mixed-priority workloads.
The first distributed LLM serving platform targeting prompt sharing. Co-optimizes KV-state reuse with computation load-balancing through a new scheduling algorithm and a hierarchical scheduling mechanism, outperforming the prior state of the art by 1.5–14.5× on average latency and 2–10× on p99.
The first LLM inference framework targeting augmented language models. Minimizes GPU resource waste caused by external interceptions and dedicates the saved memory to additional requests, improving overall serving throughput by 1.6–2× and completing 2× more requests per second over the prior state of the art.
An autotuning framework for generative-AI workflows. The AdaSeek algorithm performs hierarchical search across workflow structure, operators, and prompts under a fixed budget, improving generation quality by up to 2.8×, reducing monetary cost by up to 10×, and end-to-end latency by 2.7×.
An investigation of CPU scheduling overhead in modern LLM inference systems. The findings informed the vLLM scheduler redesign, yielding approximately 30% better performance, and were published as a widely circulated post at mlsys.wuklab.io.
A serverless scheduling framework spanning client, 5G, and cloud (AWS, GCP, Azure, Kubernetes). Heuristic and linear-programming-based code placement on compute nodes yielded 70% cost reduction and 30% performance improvement.
Investigated the effect of model size on subgroup robustness under empirical risk minimization. Counter-evidence to the hypothesis that overparameterization hurts rare-subgroup accuracy, with subgroup labels unknown.
A smart-speaker testbed studying privacy implications of always-listening devices. NLP intent classifiers and noise-filtering algorithms for passive-listening detection. Published at UC Berkeley CLTC.
A 3D Navier–Stokes water simulation in C++ with KD-tree-accelerated particle queries and parallel solvers.
A 3D-animated short about a mouse trying to reach the moon. Modeled in Maya and edited in After Effects and Premiere.