Main Page > Work Experience

Spearheaded research in Generative AI models, including advanced techniques in GANs (Generative Adversarial Networks), VAEs (Variational Autoencoders), and Diffusion models. Focused on achieving high-precision in background removal and innovating in the realm of image beautification, leveraging these models to set new standards in image processing quality and efficiency.
Integrated and customized forefront models like, Stable Diffusion. Specifically, the implementation of ControlNet played a crucial role due to its exceptional ability to manage the generative process, significantly enhancing the accuracy and versatility thus leading to superior model performance.
Employed PyTorch for developing these models. For deployment in production environments, utilized TensorRT to optimize model inference, ensuring high performance and efficiency.

Developed and accelerated 14 image/video classification and detection models using PyTorch, OpenCV, C++, CUDA, and TensorRT to identify high-risk content in online ads, focusing on sensitive topics like adult content, smoking, and vaping.
Improved model accuracy by 3.6% through experimentation with Vision-Transformer based models, notably CrossViT, surpassing the performance of traditional ResNet family models.
Enhanced user and advertiser experience by effectively filtering approximately 1.8 million inappropriate images/videos monthly with these implemented models.

Improved traffic sign detection algorithm precision by 2.1% and recall by 3.6% through innovative attention mechanisms, residual connections, atrous convolutions, and optimizers, using PyTorch and OpenCV. Enhanced model training with Albumentations for data augmentation.
Worked in collaboration with cross-functional teams to incorporate advanced localization and classification models into the internal RetinaNet framework, enhancing its detection accuracy and robustness. Successfully deployed these models in the company's car as a ROS detection node.

Developed a mobile traffic analysis system encompassing SS7, RAN, CS parts, and LTE protocols using C++ by utilizing Data Analysis to analyze errors, subscriber ID, network address, element ID, time metrics, and protocol-specific metrics to pinpoint root causes of network problems.
Enhanced processing speed by implementing DR strategies to execute 20-40 metrics simultaneously. Achieved this through extensive parallelization and optimization using SIMD vectorization, particularly AVX sets, tailored for Intel processors. Utilized various data structures and the Standard Template Library (STL) to facilitate real-time aggregation and analysis of statistical data, scaling from hundreds of thousands to millions of events per second.
This innovation significantly boosted the quality of service and user experience for mobile carriers, enabling more comprehensive network performance monitoring, including aspects like congestion, latency, packet loss, jitter, and throughput.