MdMahedi Hasan

About

Developed and deployed CaptionFace, a production vision-language system that boosts low-resolution/surveillance face recognition by generating discriminative natural-language captions (ViT encoder + GPT-2 decoder) and enabling text-to-face retrieval and zero-shot recognition. Orchestrated distributed training on Kubernetes with MLflow tracking, mixed-precision optimization, and comprehensive evaluation including out-of-domain robustness; collaborated with non-technical NSF project stakeholders via demos, visualization, and clear documentation.

Experience

Education

Key Strengths

• Built and deployed production VLM (CaptionFace) for low-resolution/surveillance face recognition
• Designed end-to-end multimodal pipeline (ViT encoder + GPT-2 decoder) for discriminative caption generation
• Improved retrieval and zero-shot recognition via cross-modal contrastive objectives and reward-based caption optimization
• Orchestrated distributed multimodal training on Kubernetes with separate pods for pipeline stages
• Operational ML skills: MLflow experiment tracking, resource tuning, mixed-precision training, monitoring via metrics/logs
• Evaluation rigor: fine-tuned vs zero-shot testing and out-of-domain robustness assessment
• Effective collaboration with non-technical stakeholders using real-time demos, visualization, and clear documentation