Ben Akhovan

End-to-end ownership of GPU computer vision frameworks, orchestration tooling, and applied ML pipelines for defense and research workloads.

• Designed and built the Universal CV family: Universal Labeler (unified detection schema, 15+ SOTA models, SAHI+TTA, VLLM auto-classification), Universal Stabilizer (streaming video stabilization DAG), Universal Tracker (frame-by-frame multi-object tracking with commit-stage semantics), and UniversalGUI (manifest-driven ReactFlow orchestration layer with FastAPI microservices).
• Engineered a scalable PyTorch training framework for classification, regression, and re-identification with CSV-labeled datasets, augmentation, class balancing, metric learning losses, and distributed multi-GPU training.
• Researched and implemented a deep learning deblurring and video enhancement pipeline with multi-GPU acceleration for high-resolution inputs.
• Contributed to vehicle counting R&D: optimized GPU/CPU pipeline usage, expanded datasets, trained and deployed TensorRT models, and delivered a client-facing web analytics tool.
• Developed fine-grained aerial semantic segmentation using deep feature and classical clustering approaches.

Built ML infrastructure for accelerometer signal analysis: training frameworks, deployment services, and performance-focused server rewrites.

• Developed a framework for training and deploying ANN models on accelerometer data.
• Built a Windows service exposing a web API for model inference, training, and monitoring.
• Used Google Colab for training and Weights & Biases for experiment tracking and performance analysis.
• Rewrote existing server functionality to improve performance and reliability; extended website and service features.