ASTRO

——————–2024——————–

Algorithm:

1. Model Training Platform Algorithms:

• Model Training & Evaluation Module: Designed and implemented a universal model training and evaluation framework supporting rapid integration and assessment of multiple task types
• Data Augmentation Module: Developed efficient data augmentation pipelines to enhance model robustness and generalization capabilities
• Video Prediction Module: Designed, developed, and integrated video prediction modules to enrich platform functionality and improve user experience

System:

1. Intelligent Annotation Software Development: Developed annotation platforms supporting automatic annotation and large model-assisted annotation, focusing on user experience and improving data annotation efficiency
2. Deep Learning Model Training Platform: Led training platform optimization involving model management, training data visualization, model inference and other functions

• PaaS / Local Training Scheduling System: Implemented multi-tenant resource isolation and training task scheduling to improve GPU resource utilization
• PaaS Resource Monitoring System: Built resource monitoring and alerting systems for real-time monitoring and visualization of key indicators including GPU, memory, and storage
• Data Storage Structure Optimization: Optimized data access structures to improve I/O efficiency during training phases and reduce storage pressure and data preprocessing bottlenecks
• Algorithm Runtime Environment Standardization: Built unified algorithm image systems supporting CPU/GPU dual-mode deployment and version management, significantly optimizing image sizes
• Training Scheduling & Model Conversion Workflow System: Built integrated workflow systems from model training and conversion to deployment, supporting multi-format model export and seamless integration with inference engines

——————–2025——————–

Algorithm:

1. Inference Engine Systems: Algorithm implementation for multi-scenario models including image classification, (rotated) object detection, instance segmentation, pose recognition, face detection, license plate recognition, including multi-model pipeline data flow design
2. Algorithm Integration: Completed full YOLO series algorithm integration supporting multi-task scenarios including object detection, instance segmentation, and rotated object detection, unifying model training, conversion, and deployment workflows
3. Model Conversion & Optimization: Implemented automated post-training model format conversion workflows (ONNX, TensorRT) to optimize inference performance
4. Data Pre-annotation Services: Designed and developed universal data pre-annotation services supporting multi-task and multi-model scenarios
5. Visual Language Model R&D and Backend Services

System:

1. Intelligent Annotation System: Led intelligent annotation system architecture reconstruction, providing object storage support, implementing K3s containerized deployment, supporting image classification, OCR, and large model pre-annotation capabilities
2. Inference Engine System: Built standardized inference plugin systems supporting multi-scenario model and algorithm deployment including image classification, (rotated) object detection, instance segmentation, pose recognition, face detection, license plate recognition, implementing dynamic loading and unloading of model services

Private PaaS Platform & GPU Resource Scheduling System (Major Automotive Enterprise)

1. Private PaaS Platform Development: Participated in overall architecture design and implementation of enterprise-specific private PaaS platforms, focusing on stability, security, and multi-user resource isolation
2. GPU Intelligent Resource Scheduling Platform: Designed and implemented distributed scalable GPU resource scheduling systems supporting full lifecycle management of one-time tasks/persistent services, intelligent resource allocation and retry mechanisms with resource observability capabilities
3. High Availability Architecture Design: Built high availability architecture based on Keepalived + DRBD, implementing automatic primary-backup switching, data synchronization, and failure recovery mechanisms to ensure core service continuity
4. Monitoring & Alerting System Construction: Deployed Prometheus + Exporters + Grafana + AlertManager full-chain monitoring systems and Loki + Promtail log stack, supporting multi-system multi-dimensional alerting strategies and intelligent grouped message pushing, covering key indicators including GPU, memory, and network monitoring
5. OTA Upgrade System: Developed OTA upgrade systems supporting remote device upgrades and version rollbacks, implementing automated device service management
6. Platform Integration & Operations Support: Coordinated with team business requirements, promoted deep integration of platforms with model training and inference workflows while supporting daily platform operations and optimization

Project Management:

1. Project Management: Responsible for project bidding, technical solution design and review for multiple modules, cross-departmental communication and coordination, driving project management from requirements to implementation
2. Code Version Management & Release: Responsible for service binary building and packaging into deb packages, reducing environmental dependencies and operational costs
3. CI/CD Pipeline Optimization: Integrated Jenkins + Ansible automated deployment, building fully automated workflows from code commits to production deployment, simplifying operational procedures

——————–2022——————–

Algorithm:

1. Posture recognition/gesture recognition algorithm research, writing, training, tuning
2. Standard posture determination algorithm development
3. RNN visualization teaching system – RNN visualization design, BPTT algorithm support & verification, question writing
4. Laboratory test sheets OCR content recognition, structured extraction

——————–2023——————–

Algorithm:

1. Delta robot arm solution simulation & verification
2. Static gesture/posture recognition algorithm and action sequence recognition algorithm design and development based on 3D skeleton recognition, and backend service development and deployment
3. Posture and gesture interaction system development
4. Speech recognition/semantic matching algorithm development and backend service development and deployment
5. Medical LLM deployment, data set collection
6. SDK design and development – speech recognition/gesture recognition/posture recognition, etc.
7. Femtosecond preoperative training system – iris recognition and scoring algorithm
8. CT image segmentation, annotation and 3D reconstruction, automatic segmentation algorithm development
9. Binocular recognition, fusion calibration algorithm development
10. Infrared small target detection algorithm development

Tools:

1. Video merging tool (Based on FFMpeg)
2. CPR posture process recognition desktop software development

——————–2024——————–

Algorithm:

1. Design and develop common CV algorithm libraries (multi-eye fusion algorithms, video, camera function algorithms)
2. Design and develop an end2end onnx model inference algorithm libraries that integrate various deep learning algorithms (advanced API, with common SOTA deep learning algorithms, fast testing, integration, and deployment)
3. Design and develop a general CS architecture service deployment framework including websocket/http (including flexible code forms and fast Docker deployment capabilities)

Project Management:

1. Write CI/CD configurations for Python package libraries, toolchains, and GUI projects

Tools:

1. Design and develop a posture editor for motion posture analysis, including a skeleton point detection tool, a point tracking tool, and a motion segmentation tool based on posture recognition algorithms

Patent & soft:

1. CPR posture process recognition patent and software copyright

Model and test related work:

• Convert various deep learning models trained based on mainstream frameworks (Tensorflow, PyTorch, Caffe, MxNet, etc.), automate conversion and data consistency testing. Form an inference model (ONNX), test data consistency before and after conversion, track problems for models that cannot be converted or have inconsistent data, and optimize the original model
• Test the inference performance and accuracy of the model on various hardware

ONNX tool chains related work:

• Maintain ONNX related tool chains, including model creation, optimization, splitting, and weight visualization
• Develop various automated tools to simplify model conversion and testing processes

Algorithm adaptation, model reproduction and demo related work:

• Maintain the self-developed model inference framework, add and upgrade various operators (the framework has CPU/GPU (CUDA)/MLU (Cambrian Deep Learning Computing Card) reasoning hardware acceleration capabilities)
• Add data pre-processing/post-processing algorithms to the self-developed model inference framework, and provide algorithm interfaces for subsequent development
• Test various SOTA deep learning models, select relatively new models for conversion and form preliminary reasoning demos (classification, detection, segmentation, migration, etc.)
• Investigate related technologies for autonomous driving, and conduct statistics on related models and feasible solutions, and form demos 

Scholarships:

• 2017-2018 Second-class Scholarship
• 2018-2019 First-class scholarship/Outstanding student leaders

Participation:

• School “Challenge Cup” College Students’ Extracurricular Academic Science and Technology Works Competition Innovation Award
• School “Internet +” College Students’ Innovation and Entrepreneurship Competition Gold Award
• Jiangsu Province “Internet +” College Students’ Innovation and Entrepreneurship Competition First Prize
• Jiangsu Province Electronic Design Competition Second Prize
• National College Students FPGA Innovation Design Invitational Competition Second Prize, Enterprise Special Award
• The third prize in the China Science Core Cup University Student Internet of Things Application Design Invitational Competition