Database & AI Systems Researcher
Final-year Ph.D. candidate in Computer Science and Engineering at the University of Michigan
Advised by Barzan Mozafari · Database Research Group
Microsoft Applied Sciences Group · Research Intern, Summer 2026
I build efficient and reliable AI & data systems. My research sits at the intersection of database systems, AI systems, and automated optimization. Across my work, I combine large language models, program synthesis, execution feedback, measurement, and verification to explore larger optimization spaces while keeping the resulting systems reliable—from SQL queries and data pipelines to long-context language models.
Before Michigan, I completed my M.Phil. and B.S. at The Chinese University of Hong Kong, where I worked on approximate maximum inner-product search.
Ph.D. expected August 2026 · Available full-time October 29, 2026 · Seeking Research Engineer, Applied Scientist, and systems roles
A research program around expanding the optimization space while keeping learned or synthesized decisions measurable and reliable.
Dependency-aware rewriting for dbt-style SQL pipelines · under review
Modern analytics are DAGs of dependent SQL transformations, but optimizers usually reason about one query at a time. DAGSmith optimizes the whole pipeline, combining dependency-aware targeting and LLM-guided rewriting with materialization and scheduling decisions.
Query rewriting via large language models · SIGMOD 2026
GenRewrite lets LLMs propose whole-query transformations beyond predefined rewrite rules, then places them inside a feedback-driven system with reusable natural-language optimization knowledge, counterexample-guided repair, and correctness/performance gates.
Whole-query optimization using program synthesis · VLDB 2023
Instead of applying a fixed library of rewrite rules, SlabCity directly searches for equivalent, faster SQL programs. Query dataflows guide synthesis, while testing, counterexamples, and verification prune incorrect candidates.
Optical context compression for long-context language models · Microsoft ASG, 2026
At Microsoft Applied Sciences Group, I studied how to carry long documents through compact visual representations rather than expensive text tokens. The project focused on learnable rendering, failure analysis, and separating reading from retrieval under aggressive compression.
Project details shown at a high level; no internal materials are linked.
* Equal contribution.