Jinchao Huang's profile picture

Jinchao Huang

About Me

I am a Ph.D. candidate in the Database Group at The Chinese University of Hong Kong (CUHK), advised by Prof. Sibo Wang. I received my B.Eng. from the University of Science and Technology of China (USTC) in 2023. I was a visiting student in the Data Systems Group at the University of Waterloo (UWaterloo) with Prof. Xiao Hu from July to December 2025, and had a short visit to the Database Group at UW–Madison, hosted by Prof. Paris Koutris in May 2026.

I work on data management for AI agents: principled methods for fast, scalable access to data. I also use agents heavily in my daily life, and many of my ideas come from watching where they struggle with data.

I am on the job market and actively looking for researcher opportunities!

Research

AI agents are becoming the heaviest consumers of data. They query at machine speed, explore speculatively, and see everything through a limited context window. While exploring, they often do not need every row of a result: a representative sample or a good estimate, delivered fast, is enough to decide the next step.

The consumer is new, but the problem is one I have worked on throughout my Ph.D.: sampling and approximate counting over results too large to compute in full.

Small answers from huge joins

Sampling and approximate counting over joins

I design algorithms that sample from a join's result without computing the join. Real queries often keep only some of the attributes, so many join results collapse into the same answer, which a uniform sample must count only once. For three fundamental classes of such join-project queries (matrix, star, and chain), I gave the first asymptotically optimal algorithms for uniform sampling (PODS'26). Queries also filter: for acyclic joins under equality filters supplied at query time, a disjunction of filters makes the result a union of overlapping sets, and my union sampling algorithm draws each uniform sample in expected time linear in the number of sets, which is optimal (ICDT'26). I also brought subset sampling to joins: each join result is kept independently with its own probability (PODS'26).

Samples that keep up with change

Dynamic subset sampling

Agents work on data that changes under them: records arrive, scores are revised, memories are added and re-weighted. My indexes keep samples correct as the data changes, without rebuilding: Poisson πps sampling under updates (KDD'25) and subset sampling over joins under insertions (PODS'26).

Spending a limited budget well

Budgeted data selection

Filling a fixed context window with the most useful data is a budgeted selection problem. I have studied such a problem in another setting: choosing which nodes to label, under a fixed labeling budget, when training graph neural networks (VLDBJ'24).

Next: data systems built for agents

I want to build on this work toward systems whose answers fit an agent's token budget and accuracy needs, and that share computation across the many overlapping queries agents issue.

Publications

α alphabetical order (theory papers convention) # equal contribution

Selected Awards

Services

Teaching