Omid Aghdaei

Developed production ML systems for energy forecasting using statistical, deep learning, ensemble, and probabilistic methods. Built automated time-series evaluation workflows and metadata-driven ETL pipelines. Implemented an MLOps platform with Airflow, MLflow, and Streamlit for automated training, tracking, and evaluation. Optimized Oracle SQL/PLSQL for large-scale data processing. Mentored ML engineers in software architecture, pipeline design, and code quality.

Led the design and production deployment of real-time Persian ASR for telephony, improving accuracy in noisy environments; contributed to microphone-based ASR. Optimized low-latency inference using distillation, quantization, ONNX Runtime, and Flask. Developed streaming speaker diarization and VAD workflows with pyannote. Deployed a real-time pronunciation assessment engine and standardized data curation, hyperparameter tuning, and benchmarking for production readiness.

Developed an autonomous drone navigation framework using D3QN and ResNet-8 in AirSim, improving obstacle avoidance by 2% and inference speed by 18%. Trained vision models achieving 95.4% accuracy at 20 FPS. Integrated MiDaS depth estimation for improved obstacle perception. Optimized inference with ONNX Runtime, pruning, and quantization, boosting CPU speed by 25% with only 1.4% accuracy loss. Built a YOLO-based hazard detection pipeline using manually annotated data.

Thesis title: Quadcopter Autonomous Navigation using Deep Reinforcement Learning Grade: 3.61/4.00 GPA

Thesis title: The relation of Internet of Things in smart city and smart car Grade: 3.31/4.00 GPA