WENSHENG (WINSON) WU
Full-Stack AI Engineer | Master of Artificial Intelligence (Candidate)
📍 Christchurch, New Zealand | ✉️ wensheng.wu@pg.canterbury.ac.nz
EDUCATION
Master of Artificial Intelligence (Candidate)
University of Canterbury | Feb 2025 – Jul 2026
- Relevant Coursework: Machine Learning, Deep Learning, Computer Vision, AI Ethics
- Focus Areas: LLM Applications, Edge AI
Bachelor’s Degree
Dongguan University of Technology | Sept 2008 – Jun 2012
- Major: Communication Engineering
EMPLOYMENT
Architect & Software Engineer
Intellifusion (www.intellif.com) | Aug 2018 – Sep 2024 A leading Chinese AI company (IPO in 2023, NZ $8 billion market cap)
- Recognized as a Top 20% OKR performer, delivering consistent high-impact results.
- Designed and implemented scalable system architectures and core modules, ensuring high availability and performance for AI-driven platforms.
- Improved team-wide code quality and maintainability by leading rigorous reviews and establishing coding standards and best practices.
- Shaped the technical roadmap, driving AI model integration and adoption of emerging technologies, while delivering production-grade applications and published 9 technical patents.
Software Engineer
Jobcn (www.jobcn.com) | Jan 2017 – Aug 2018 Employment website for job listings
- Rebuilt the job and resume search engine, improving query performance and scalability, reducing average query latency by 50% and supporting 2.5× higher query throughput.
Software Engineer
ChinaRS (www.rscloudmart.com) | Apr 2016 – Jan 2017 Remote sensing data workstation
- Developed data management modules for remote sensing, supporting standard and custom workflows.
IT Support Engineer
Dongguan Civil Affairs Bureau (www.mzj.dg.gov.cn) | Sept 2012 – Apr 2016 Government agency for social administration
- Managed IT infrastructure, ensuring secure and reliable operation of servers, networks, and security devices.
SELECTED PROJECTS
Large Language Model Development (YunTianshu, up to 70B parameters)
Ranked #1 on the C-Eval Chinese LLM leaderboard (Aug 2023) with an average score of 77.1, surpassing 62 competing models including GPT-4. Role: Architect & Developer | Jan 2023 – Sep 2024
- Led the development of 6B, 13B, and 70B LLMs, trained on a 24TB raw multilingual corpus.
- Designed and implemented distributed training architecture, enabling efficient training on on large GPU clusters.
- Optimized the training and fine-tuning pipeline, covering corpus collection, data preprocessing, model training, model optimization, and alignment, improving training efficiency and stability.
AI Agent Service Platform
An enterprise AI Assistants platform powered by LLMs, designed for AI capability development and integration across internal business systems. Role: Architect & Backend Engineer | Jun 2023 – Sep 2024
- Architected and implemented a scalable AI assistant platform, delivering core services including conversation management, structured function calling, document retrieval, hybrid memory, security controls, prompt templating, and dynamic tool routing.
- Built a unified abstraction layer to integrate OpenAI GPT-4 and proprietary LLMs, enabling dynamic model routing and standardized function invocation.
- Enabled deployment of 10+ internal AI-powered products across multiple departments, accelerating enterprise-wide adoption of LLM-driven automation.
Large-Scale Face Image Retrieval & Big Data Platform (DeepEye)
One of the world’s largest face retrieval systems, enabling real-time search across tens of billions of images within seconds on a large-scale GPU cluster, and ranked Top 3 in accuracy nationwide (China). Role: Architect & Backend Engineer | Aug 2018 – Jan 2023
- Scaled core services by migrating services to Kubernetes, expanding the cluster from 20 to 370 nodes, and boosting throughput to 14.3 PFLOPS, ensuring elastic scalability under rapidly growing data demands.
- Redesigned face retrieval engine, reducing query latency and increasing single-server capacity from 50M to 500M indexed faces.
- Built CUDA-accelerated applications for large-scale feature extraction, retrieval, and video decoding, achieving 18,000 image feature computations per second.
YMIR (https://github.com/IndustryEssentials/ymir)
A Rapid Data-centric Development Platform for Vision Applications open-sourced in 2021, and was awarded the “2021 Wu Wenjun AI Science and Technology Award,” the highest accolade in China’s AI sector. Role: Backend Engineer | Aug 2020 – Nov 20221
- Designed and developed core code of the Iteration, Dataset, Model and Algorithm Management modules.
- Resolved technical challenges in model training related to version rollback and cross-version compatibility of annotated data.
- Participated in drafting the technical documentation, which has assisted 20+ companies in producing their own industrial models.
PATENTS
- Face image normalization method (CN117876296, 2023) – Improved face recognition accuracy for computer vision systems.
- Image detection area determination method (CN116168192, 2022) – Customized camera detection zones using Tenengrad function with Sobel operator.
- River channel monitoring method (CN115861932, 2022) – Enabled multi-source image-based pollutant tracking for river water quality monitoring.
- People flow detection method (CN114373198, 2021) – Enhanced real-time crowd counting accuracy in closed environments using facial feature clustering.
- Signal lamp control method (CN113936476, 2021) – Optimized traffic flow through adaptive signal control based on vehicle speed and density.
- Equipment clustering method (CN113901981, 2021) – Automated scene grouping using Laplacian matrix and spectral clustering techniques.
- Job matching method (CN114693231, 2020) – Improved occupation matching accuracy via similarity computation of job feature libraries.
- Hotspot route analysis method (CN114637883, 2020) – Predicted individual movement patterns using public transportation data for safety administration.
- Human body thermodynamic chart methods of exhibiting (CN109816745, 2018) – Generated thermal heatmaps through segmentation of detected human bounding boxes.
SELECTED ACHIEVEMENTS
Most Popular Award | Startup Weekend (Dongguan, China) | Dec 2016
- Startup Weekend, powered by Google for Entrepreneurs, is non-profit entrepreneurship education event.
- Role: Team leader & Full-Stack Developer
- Led a 7-member team to propose and develop a startup concept, conducted market research, built MVP, and pitched to judges and investors.
SKILLS
- Programming: Java, SQL, Python, C++, Shell
- AI & ML: TensorFlow, PyTorch, TFLite, Keras, FAISS, BGE, LangChain, Dify, AutoGPT, CUDA
- Data & Storage: MySQL, PostgreSQL, Oracle, SQLite, MongoDB, Redis, Elasticsearch, Apache solr
- Infrastructure & DevOps: Linux, Docker, Kubernetes, Git, Nginx, Tomcat
- Distributed Systems: Kafka, Spark, Azkaban