I’m Qiaodan Ju

Huawei Technologies Co., Ltd

Wireless Communications Dept

Software Development Engineer

Lead the design and development of multiple failure mode embedded plug-in modularization for automated test factory, using AI to iteratively output non-product problem confidence based on historical data and terminal and base station characteristics, and sinking into product test line as well as solution test process to improve test efficiency by 5-10 man-days/month.

Independently undertook the design and development of the front and back end of the visualization portal and its database for the defense line factory, and was awarded the institute level QCC.

Participated in use case analysis and line of defense construction for each iteration of millimeter wave products, with 0 product issue outflows, and sunk multiple general efficiency tools.

Terminal Dept

Software Development Engineer

Led the animation system interaction between U3 engine and UE, lightly customized UE5 engine to reach compatibility with one-click resource import on U3 side.

Participate in XXXX and XXX and other Tencent headline games (Unity engine), such as engine hongmongization, client hongmongization and pipeline construction and other adaptation work.

Master of Science in Data Science and Machine Learning

GPA: 4.33 / 5.00 (93.3 / 100.0),

Bachelor of Engineering in Software Engineering, Southeastern University, China

GPA: 3.9199 / 5.0000 (89.199 / 100.000),

CET-6: 610

IELTS: 7.0

GRE: 331 + 3.5

Megvii-inc, Beijing, China

Working on a 3D lane detection model to build an easy-to-deploy network that works efficiently end-to-end. Used ResNeXt18 as the backbone and connected the backbone and the neck of the dcn_fpn structure via an SA module to enhance global feature awareness.

Participating in the improvement of the prediction head, the clustering strategy of key-points in post-processing is designed to draw lane lines under RV segmentally with different joint_points as bridging points. The lane lines are merged by certain strategies including IoU, offset, etc. to give the model robustness in the diverging lane lines and thus replace the unfriendly RIM solution deployed in condlanenet.

Write data pre-processing and GT annotation pipeline for curvelane, CULane, etc. Distributed training based on MMCV framework, as well as test scripts and asynchronous upload of bad cases to s3 for visual analysis and checkpoints via a mounted ws2 multi-thread. Versioning.

NUS Next++ Lab

Involved in the research of the Image Caption task network model, mainly in the backbone refactoring optimization, the algorithm design of the caption evaluation and the redesign of an adaptive and efficient optimizer.

In the Backbone module, Inception-V3 was improved with reference to DLA34's tree aggregation idea to provide more global awareness and maintain a healthy number of participants. The evaluation method of the original Greedy Search is replaced by the Beam Search strategy, which generates a more biased prediction result in favour of selecting the best.

An optimiser between Lazy Adam and Adam is redesigned and combined with AdamW for regular term tuning. The problem of Lazy Adam not updating first- and second-order momentum at zero gradient is addressed by making it applicable to the case of sparse variable indexing during RNN training, and for unsampled embedding plus penalty coefficients is also taken into account. After passing through the neck of dcn_fpn it is fed into the RNN head with RA module, using the GRU cell as a carrier. In a preliminary demo using the VizWiz dataset, its BLEU_1 reaches 50 and BLEU_n can basically reach around 30.

Neusoft Technology Co., China

Please refer to the link for details of the project open source code: https://github.com/shimmerjordan/sz-education

The system is designed to improve the efficiency of online teaching and learning and includes an online learning sub-system and a back-end CMS, which contains modules related to online courses, online examinations, grade analysis and grade management, intelligent paper assembly, tenant management and other related functions. The system uses an improved genetic algorithm to improve usability and scientific performance.

The system framework is based on SpringCloud and integrates monitoring components such as Consul, Spring Boot Admin and Zipkin. Spring Cloud OAuth2 and JWT are used to support multi-tenant SSO login and authentication, and RabbitMQ is used to maintain the message queue. In addition, BCryptPasswordEncoder method is used for irreversible password encryption. Combined with Redis for partial request caching and filtering, the system was tested to maintain a concurrency of over 10kQPS when deployed locally in a package.

The system was packaged and deployed on a local Ubuntu 20.04 test machine and deployed to the public network via intranet penetration.

Chinasoft International, China

For details, please refer to the link. https://github.com/shimmerjordan/Lexian-Malls

Participated in the development of Lexian-Mall supermarket fresh food wholesale system based on SpringCloud, including client, applet and CMS backend management system. Mainly responsible for the SSO module for users under the login authentication microservice, the real name authentication function module and the shopping cart management module under the order microservice, and the development of some functions under the order module.

Developed the client and CMS system web front-end and applet pages based on Uni-app and Vue respectively, and used SpringBoot to develop each functional module. We used Redis to solve the problem of unshared sessions and implemented a simple message queue to speed up the response to requests and improve concurrency performance. In addition, Zuul was used to intercept and filter illegal requests from requesting users to intercept malicious attempts at cryptography.

The application was packaged and deployed to a cloud server after JVM tuning and passed a 10kQPS stress test in a test environment.

User Debt Behaviour Forecasting - kaggle competition

For details, please refer to the link. https://github.com/shimmerjordan/loan-pred-customer-behavior

This (unbalanced) dataset is studied from supervised as well as unsupervised learning directions, focusing on feature engineering, modelling prediction and dimensionality reduction, and user clustering respectively. The main objective is to build highly accurate predictive models of user lending behaviour.

The XGClassifier was used to train the optimal ratio of undersampling and oversampling to better address the data imbalance. Feature selection was performed by PCA and Variational Auto-encoder, and a pipeline was established for GB, RF and Xgboost respectively, and the Xgb model with a training set of 0.97 f1-score and a test set of 0.89 f1-score was obtained by model selection. In addition, k-means was used to classify the users for model interpretation.

Automated Palletizing System

Description: Realized machine’s automatic identification to assess the box’s height and size, placed the messy boxes in object region orderly using mechanical arms.

The coding of the movement logic of the robot arm of the automatic palletizing system allows the arm to send requests and responses to the corresponding algorithm module after it has obtained the commands to achieve automatic palletizing and loading.

Image Identification Based Upon Progressive Machine Learning

Description: Applied progressive identification techniques to the image classification problems of bags and suitcases, increased the efficiency of image classification in case of ensuring accuracy.

Wu Jianxing, Zeng Dexin, Ju Qiaodan, Chang Zixuan and Yu Hai*, “Progressive Image Recognition Method and Its Application in Security Inspection Machines”, Current Chinese Computer Science 2021; 1(1) . https://doi.org/10.2174/2665997201999201208210819

Participate in the design of classification networks and the construction of datasets.

The initial dataset is obtained through a crawler script on an e-commerce platform, and then combined with manual GT annotation through data cleaning and augmentation and other pipelines. A progressive strategy is used to segment individual images to simulate the form of progressive input.

The VGG16 is optimised with reference to the BN proposed by InceptionV2 and the ResNet residual linking, and the CategoricalCrossentropy is improved by combining the focal loss proposed by RetinaNet, for different proportions of segmented data from the same image source, the small proportion of misclassified cut data will have a larger penalty factor, as a way to control the convergence efficiency of the model.

Ultimately, with only 50% of the images input, it is able to obtain a 50% efficiency gain at a cost of less than 1.5% precision loss (precision == 90.35%).

Study on Electricity Generating Performance by Burning RDF

Description: Developed research on RDF’s composition proportion under optimal burning performance.

Analysis of the current status of research and directions for improvement in the performance of RDF fuels for power generation.

The coarse material of rdf is made into combustible materials with different mixture ratios by controlled variable method.

Based on experimental statistics, the combustion characteristics of rdf blends are analysed in terms of weight loss temperature and ash analysis.

Undergraduate Level

Second-class Scholarship for Excellent Students, NEU, China. 06/2021

Successful Participant, 2020 Certificate Authority Cup InternationalMATHEMATICAL CONTEST IN MODELING. 05/2021

Outstanding Student Pioneer at NEU, China. 04/2021

Second-class Scholarship for Excellent Students, NEU, China. 09/2020

Outstanding Student at NEU, China. 11/2019

Third-class Scholarship for Excellent Students, NEU, China. 10/2019

Third Prize at university level, The 5th “Internet +” Entrepreneurship and Innovation Competition, China. 06/2019

Successful Participant, Mathematical Contest in Modeling, the US. 05/2019

Outstanding Minster, Academy of Science and Technology, NEU, China . 04/2019

First-class YU Songling Scholarship, NEU, China. 11/2018

Third Prize at provincial level, National Mathematical Contest in Modeling for College Students, China. 10/2018

Second-class Scholarship for Excellent Students, NEU, China. 09/2018

Second Prize, Mathematical Contest for College Students, NEU, China. 07/2018

Seventh place in the men's single-shake jump rope category of the 57th Tohoku University Games

Second Prize, “PUYI” National English Composition Contest for College Students, China. 12/2017

2022 NUS Distinction Honour Graduates

2023 Huawei NEO Distinction Honoree; DPC Distinction Camper; Excellent Turnaround Intern; Star of Tomorrow