At Khorasan Science & Technology Park, I research emerging AI developments, publish technical and educational content, and maintain Hooshub.com. I also customize WordPress components and make technical improvements to the platform.
At ParsTech.Co, I designed and trained Persian RoBERTa masked language models using more than 1.2 TB of collected Persian text. I built distributed training pipelines across multiple servers and GPUs, and developed a Persian semantic tokenizer that reduced vocabulary size by up to 80% compared with BPE.
My tokenizer and training improvements raised MLM model accuracy by 20%. I also led NLP solution development and mentored four coworkers in NLP and machine learning.
Earlier at ParsTech.Co, I trained YOLOv5 on an Iranian license plate dataset, achieving 98% accuracy, and fine-tuned EasyOCR for Iranian identity cards, achieving 99% accuracy. I built a license plate generator and Python tooling for data preparation, augmentation, training, and evaluation.
