Karam Mahfod

Engineered and deployed a real-time AI pipeline for attendance tracking, capable of simultaneously processing over 300 4K camera streams, significantly reducing GPU load through ROI-based detection while maintaining high recognition accuracy.

Developed a multi-model pipeline leveraging pose estimation and action classification to recognize various student behaviors and analyze teacher-student interaction patterns in real-time, providing actionable insights for educational improvement.

Spearheaded platform expansion initiatives, integrating voice-based session summaries and advanced multi-camera tracking, which enhanced inference performance for large-scale deployments.

Led a team of four engineers in the research and development of next-generation neural image compression methods, contributing directly to the JPEG-AI standardization effort.

Developed a novel perceptual quality metric, trained on Mean Opinion Score (MOS) and validated against subjective human evaluations, achieving superior correlation with human perception compared to traditional metrics (VMAF, MS-SSIM, PSNR).

Delivered significant gains in compression efficiency without perceptual quality loss, achieving 25% bitrate reduction for the 0.075 bpp model and 31% for the 0.012 bpp model compared to baseline methods.

Optimized fixed-quality and fixed-bitrate training paradigms, balancing efficiency and visual fidelity across diverse model sizes for enhanced image compression.

Optimized H.265/HEVC video codecs, enhancing compression efficiency and runtime performance across multiple platforms by redesigning encoder/decoder modules and optimizing rate-distortion, motion estimation, and entropy coding.

Implemented advanced multithreading strategies and low-level optimizations, boosting throughput and scalability, and significantly improving pipeline concurrency and memory utilization in real-time and batch processing.

Developed a feature-rich 32-bit RISC-V simulator, incorporating multithreading, MESI cache coherence, and branch prediction for hardware-software co-design and codec acceleration.

Created custom profiling and debugging utilities utilizing GProf, Valgrind, and low-level instrumentation to pinpoint bottlenecks and validate system-level performance improvements.

Contributed to CI/CD workflows, cross-platform builds, and automated testing pipelines, leveraging containerized environments and continuous benchmarking for robust video compression solutions.

Designed and implemented a comprehensive transcoding pipeline utilizing OpenHEVC and x265, enabling efficient video re-encoding and format conversion for large-scale video processing workflows.

Integrated SUR-based quantization and HVS-inspired pre-filtering to optimize quantization decisions, significantly improving both compression performance and visual quality.

Re-architected decision logic to shift computationally expensive processes from decoder to encoder, significantly reducing overall encoding complexity and runtime.

Achieved equal or superior compression efficiency compared to H.266 for the enhanced H.265/HEVC pipeline across multiple video datasets, validated through BD-Rate analysis.

Developed robust backend C++ modules for real-time telecom network management systems, ensuring high availability and performance.

Collaborated effectively with cross-functional teams to deliver new features and resolve complex issues in a fast-paced development environment.

Gained foundational experience in large-scale system development and maintenance, contributing to critical infrastructure projects post-graduation.