Kirill Molchanov

• Optimized real-time models for 3x faster processing (TensorRT), reduced client CPU load (19%→11%), & enabled 30 FPS mobile gaze prediction. • Boosted model accuracy, increasing forbidden object detection F1 by 20% & graphical object detection F1 from 76%→85%. • Enhanced Super Resolution models (83% user preference) & improved QA models by cutting False Negatives by 21%. • Replaced legacy algorithms with modern DL solutions (e.g., MobileNet) & cut feature development time from weeks to days.

Project - Secure Perimeter, Secure Workspace Monitoring: • Migrated from classical computer vision algorithms to modern architectures like MobileNet for people detection, balancing safety and performance on edge devices. • Optimized and deployed deep learning models for real-time applications across client, server, and mobile platforms, increased frame processing speed by 3 times (using TensorRT), reduced client-side CPU load (19% → 11%), and enabled 30 FPS gaze prediction on mobile.

Projects: Intel Graphics Performance Analyzers, OpenVino VPU plugin • Developed performance analysis features for Intel GPA, optimizing data processing algorithms to help developers diagnose bottlenecks on Intel CPUs & GPUs. • Optimized neural network inference on pre-production Intel VPUs as a key contributor to the OpenVINO plugin. • Boosted model performance & compatibility by implementing graph optimizations, quantization, advanced tiling, and low-level runtime vectorization.