CV

M.S.

Northeastern University, Khoury College of Computer Sciences

Computer Science

• Parallel Data Processing, Algorithms, Computer Networking, Cloud Computing, Database Management Systems
• GPA: 4.0/4.0
• Completed at Northeastern’s Seattle campus.
• Teaching Assistant for Algorithms and Parallel Data Processing.

M.S.

University of Southern California, Viterbi School of Engineering

Applied Data Science

• Scientific Computing and Visualization, Analysis of Algorithms, Machine Learning, Data Mining, Database Systems
• GPA: 3.83/4.0

B.S.

Wuhan University, School of Resource and Environmental Sciences

Geographical Information Science

• Data Structures, Linear Algebra, Discrete Mathematics, Probability and Mathematical Statistics, C Programming
• GPA: 3.58/4.0

Software Development Engineer

Beaconfire Solutions Inc.

• Developed enterprise HR and digital-commerce microservices with Spring Boot/Spring Cloud, integrating authenticated APIs, relational/document data stores, AWS S3, and RabbitMQ-backed asynchronous workflows.

Teaching Assistant — Algorithms and Parallel Data Processing

Northeastern University, Khoury College of Computer Sciences

• Supported graduate coursework in algorithms and distributed data processing, covering MapReduce, Spark, and parallel execution models.

R&D Engineer

NSFOCUS Technologies Group Co., Ltd.

• Built and maintained distributed security-data pipelines using Kafka and Elasticsearch for multi-source log ingestion and normalization, and implemented Spark/MapReduce batch workflows serving 100+ enterprise clients and processing approximately 5M records per day.
• Developed host-anomaly detection features for illicit file transfers and backdoor processes using fuzzy inference, hierarchical clustering, and k-core graph mining over multidimensional security logs.
• Contributed to a production alerting workflow that achieved effective response for 100% of severe threats and 90% of high-risk threats in a telecom deployment.

R&D Engineer Intern

NSFOCUS Technologies Group Co., Ltd.

• Processed and normalized multi-source security logs covering user operations, file access, and process behavior to support downstream host-anomaly analysis.

Web-Based Dynamic Thematic Mapping Technologies

Undergraduate Thesis — SinoMaps Press Collaboration

Advisors: Prof. Haihong Zhu and Yuntao Long

• Surveyed web-based thematic-mapping methods and designed the overall, functional, and logical architecture for online visualization, interaction, and dynamic data updates.
• Built and validated a D3.js customs trade-map and map-service prototype supporting thematic visualization, statistical charts, interactive queries, and live data updates.

PM2.5 Inversion and Spatiotemporal Analysis for the Wuhan Metropolitan Area — Core Team Member

National Undergraduate Innovation and Entrepreneurship Training Program

Advisor: Prof. Huanfeng Shen

• Integrated ground-monitoring measurements, NASA MODIS/AOD satellite observations, and meteorological data, performing preprocessing and spatiotemporal matching across multiple data sources.
• Contributed to the development and evaluation of a mixed-effects model for PM2.5 inversion, using cross-validation and baseline comparisons to assess predictive performance.
• Used MATLAB/R for modelling and ArcGIS to generate spatially continuous PM2.5 concentration maps across the Wuhan “1+8” metropolitan area; the national program passed final review with a Good rating.

Multi-GPU Similarity Engine — Lossless Collective Compression & Overlap

C++, CUDA, NCCL, MPI, OpenMP, PySpark, CMake, Nsight Systems — extension of USC INF 553 Data Mining / CSCI 596 Scientific Computing projects

• Re-engineered a Yelp collaborative-filtering workload from Spark into serial C++, OpenMP, MPI, CUDA, and dual-GPU NCCL backends behind a frozen numerical contract; every backend bit-exact, processing 1.17M candidate pairs in 4.21 ms on two NVIDIA A100s — 312× serial, ~20× over 16-thread OpenMP.
• Compressed the NCCL AllReduce payload 3× losslessly (56.2 → 18.8 MB) by packing six integer sufficient statistics into two uint64 words with carry-free fields, letting NCCL sum them in compressed form without decompressing; verified bit-exact across 2.58M pairs.
• Isolated the interconnect by running the identical binary on NVLink and PCIe — communication is 11% vs 91% of the iteration, so the same compression buys 4% vs 55%, and combined with overlap cuts the PCIe iteration 2.53×.
• Derived and Nsight-verified the overlap ceiling min(compute, comm)/total — overlap pays when compute and communication are comparable, not when communication is large; a chunk-size sweep also exposed and fixed a double-buffering race.

Quantized LLM Inference Study on a LoRA-Fine-Tuned Llama 3.1 8B

Python, PyTorch, CUDA, PEFT/LoRA, bitsandbytes, Hugging Face Transformers, Modal — independent inference-systems study extending an LLM-engineering course fine-tune

• Fine-tuned Meta Llama 3.1 8B with QLoRA (rank 32 / α 64 on attention projections, 27M trainable parameters — 0.34% of base) under completion-only supervision on a curated 400,000-item dataset, and served it 4-bit NF4 on an NVIDIA T4 via Modal.
• Built a CUDA-event inference benchmark across four quantization configurations on Turing and Ampere, decomposing latency into prefill and decode against an analytical roofline; showed 97% of the tokens are prefill but only 39% of the time is — decode costs 54× more per token.
• Identified a 2.50× end-to-end regression in the deployed configuration — a bf16 compute dtype carried over from Ampere training onto Turing, which has no bf16 tensor cores. Localized entirely to prefill (4.42×, MFU collapsing to 3.2%) and isolated by an unquantized control agreeing within 0.8%.
• {“Showed 4-bit quantization buys capacity, not speed — NF4 decode runs 1.69× slower than fp16 while moving 3.9× fewer bytes, reaching 9.8% of peak bandwidth against fp16’s 64.1%”=>”dequantization-bound, not bandwidth-bound.”}

Zephyr Manus — Agent Execution and Tool-Orchestration Platform

FastAPI, Redis Streams, PostgreSQL, Docker, Next.js, MCP/A2A — independent project

• Built a Planner–ReAct agent runtime supporting task decomposition, external-tool execution, persistent state, and human-in-the-loop wait/resume workflows.
• Developed Docker-isolated execution for Browser, Shell, and File tools and an event-driven FastAPI backend using Redis Streams, PostgreSQL, SSE, and MCP/A2A.

Train Ticket Flash-Sale System

Java, Spring Boot, Spring Cloud, Redis, RocketMQ, MySQL — independent project

• Built a high-concurrency ticket-purchasing service using Redis-based inventory control and RocketMQ-based asynchronous processing to decouple request handling from downstream order creation.
• Designed idempotency and consistency mechanisms for concurrent requests, mitigating duplicate purchases, overselling, and retry-related failures under flash-sale workloads.

Distributed Apriori Mining on Yelp Dataset

Java, Hadoop MapReduce, HBase, Hive, AWS S3 — Northeastern CS 6240 Parallel Data Processing

• Built a distributed Apriori pipeline on Hadoop MapReduce over ~550K Yelp reviews and ~15K businesses, decomposing candidate generation, support counting, pruning, and iterative frequent-itemset discovery, with Hive external tables preprocessing raw JSON into per-user review baskets.
• Implemented two interchangeable storage backends for cross-pass state — an HBase path materializing each pass as a table with support filtering applied at write time, and a file-based path over HDFS/S3 with a runtime flag switching between local and S3 discovery of per-pass outputs.

Outstanding Student Leader

Wuhan University

Awarded in both the 2014–2015 and 2015–2016 academic years.

Outstanding Individual in Theoretical Study

Wuhan University