Shenghao L.

Web data acquisition for LLMs & GenAI.

Co-initiated PyTorch's StateDataLoader, a data loader supporting auto checkpointing and resume. StatefulDataLoader is available in PyTorch open source, and is integrated into HuggingFace.

Designed the data placement algorithm for MAST, a global scheduling of ML training across geo-distributed datacenters at hyperscale. Paper is published in OSDI'24 (https://www.usenix.org/system/files/osdi24-choudhury.pdf).

- Formal Verification of Neural Turing Machine, with Prof. Manos Kapritsos - The Study of Social Media's Influence on Social Movements, with Prof. Ceren Budak

Core System - Tupperware Platform • Optimized Tupperware Platform metadata storage format, saving ~75% memory and ~50% network usage on Read layer • Built adaptive fields in API response, to avoid network communication overhead • Developed performance regression detection tool for Read APIs • Created CLI tools for Tupperware Platform to perform manual modification on Zeus • Built domain-lookup endpoints on sharded DBs

How Photon and Cortical Noises Limit What We See • Built human cone cell absorptions model for various environments in Matlab using iset-bio package • Constructed a supervised classifier to measure threshold energy of photon cells with different noises • Calculated equivalent noise of photon cells based on the result of their threshold energy • Assisted in human visional perception testing, helping to design a stable and reliable environment for testing • Constructed formula to describe the relation between the strength of internal equivalent noise of photon cells and the energy level of outside noise

Stats and Discrete Maths

Project: Study of Feedforward Inhibition and Synaptic Scaling in Argyle Illusion • Assisted in formula proof and verification in the algorithm, especially the approximation in weight evolution • Adjusted parameters in expectation-maximal algorithm to balance weights of prior judgement and rate of post learning • Built simulation data and run them in the model, which I managed to fit into the real-world data • Tested the case when different classes have an overlapping and managed to distinguish them accurately • Built shadow detection, specifically on introducing global light variable and fit it to current model

• Assisted in analyzing macroeconomic trends in different industries in China in the first half year of 2016 • Collected information regarding the expectation and reactions from world major companies in finical field to Brexit and further predicted future risk of the current asset due to this event • Assisted in adjusting the company’s asset allocation and strategy, moving the focus from stocks to bonds • Made risk assessment for investment projects, mainly in charge of the abnormal leverage analysis

• Assisted in the research on the personal consumption loan in China based on the recent 20-year data • Analyzed the relationship between social and political events and personal consumption loan’s development • Built the regional personal loan distributions and their implications of the different roles that different regional economic level and culture play in the growth of personal consumption loans • Made a vintage analysis for consumer finance companies’ credit score based on the time classification of their customers

• Responsible for improving features of BON APP!, a food-searching product targeting serious food lovers in Shanghai • Designed key gestures of Meet function, which event gathers stranger users and have a theme meal together • Built an exponential model constrained by partial derivatives for sorting user reviews to balance time and popularity • Applied A/B test and used non-parametric inference to evaluate users’ preference level for the new icon placement

Double Major in Mathematics and Computer Science, magna cum laude