Guowei Wang Email & Phone Number

Hangzhou, Zhejiang, China

+ Research and development on algorithms for model training / interpreting compression and accelerating.+ Analyze model accuracy and performance after porting to a new platform.+ LLM inference accelerating methods like static KV cache / quantization / tensor parallel / speculative decoding.+ [Keras cv attention models](https://github.com/leondgarse/keras_cv_attention_models) - This project is the Keras / Tensorflow implementation of most SOTA CV models, also including ported Imagenet / COCO pre-trained weights, from other frameworks like PyToch / Jax / PaddlePaddle. Currently, most latest Imagenet model Tensorflow implementations in paperwithcode come from this project. - [Recognition models on Imagenet]. Such as FastViT, LCNet, CoAtNet, Uniformer. Most models are the only implementation for Keras on GitHub. - [Detection models on COCO]. Including EfficientDet, YOLOR / YOLOX / YOLOV8. As an improvement, realized the combination of any backbone + detection header + anchor loss. - [LLM models and CLIP training]. Including GPT2, LLaMA2 model architecture and CLIP training. - Other CV model like DDPM training / Stable Diffusion / Segment Anything model implementation. - [Model training]. Contains implementation of progressive training, token label training, model distillation and other advanced methods. - [PyTorch backend]. Supports using PyTorch as backend for Keras API, means a same model definition code can be used directly on TensorFlow or PyTorch.

Qingdao, Shandong, China

+ Face recognition, PAD related algorithm R&D. Keras model optimization, and replicating other recent article results. Deployment related including converting TFLite / TensorRT models for mobile devices and server usage.+ [Keras insightface](https://github.com/leondgarse/Keras_insightface) - This project is the Keras implementation of Insightface, including basic arcface and other paper implementations like magface / curricular / triplet, also other training methods such as subcenter / VPL - [Model training] At present, EffientnetV2S using ms1mv3 with 5M data volume reached IJBC TAR@FAR 1e-4 = 0.971212. Relatively, original Insightface R50 model using Glint360k with 17M data is 0.970292. - [Model deployment] Mainly used inference frameworks are TFlite / NCNN / TensorRT. Go is used for packaging Android AAR/JAR, and providing server TensorRT interface. - [Actual scenarios usage] Solving issues like CPU/memory usage on device, camera calibration of actual hardware, face quality estimating, and optimizing entire calling process from Android -> Go -> TFLite.

Qingdao, Shandong, China

+ Response for the algorithm usage in productions, including image / text recognition and classification. Also help providing AI practice courses. - [Text semantic analysis] Built an intelligent customer service, including Jieba Chinese processing, automatic reply of AIML pattern matching mode, TF-IDF model and cosine similarity matching, as well as the simple application of seq2seq model. - [Course recommendation algorithm]. Extracting multiple user portrait dimensions to describe a user, for subsequent courses recommendation. Also developed a skill scoring system. Then applied Neo4j graph database planning user learning path.

Qingdao, Shandong, China

- Main responsibility is software R&D in SIP voice service of AONT, a household optical fiber router. Also help analysis the solution of customer needs from multiple countries, and timely answer the questions raised by customers.- The entire project is a large scale C++ one, which prefer multi person collaborative development, and team code review for maintaining code quality.- Mainly development is based on ARM chip, involves bottom driver adaptation, network communication, SIP protocol voice service, and implementing other customer scenarios.- During this period, also participates in code compilation and review process optimization, compilation server deployment, etc.

Shenzhen, Guangdong, China

- Take part in embedded medical equipment development. Including graphical user interface, digitized data waveform, alarm and other sensor hardware drivers adaption development.- GUI display is mainly using Minigui, realizing functional split, waveform plot and data display with multiple forms.- Other software related like integrating third party medical library, realizing cardiograph waveform plot, blood pressure and other abnormal status detection and alarm.