Yassien Ashraf

I’m a deep learning practitioner focused on computer vision, multimodal learning, and practical ML engineering. I contribute to Keras through the KerasHub ecosystem, where I’ve added support for BLIP-2 and reported over 10 bugs that got fixed.

Working in open source taught me how production software gets reviewed and shipped to a large user base. I regularly use Git PRs and issues to collaborate and track progress end-to-end.

In my internships and training, I’ve applied deep learning to real constraints—like data privacy. At Cegedim, I worked on an NLP project for datamasking and PII detection to meet EU GDPR requirements.

I also build and iterate on strong model prototypes: TransU-NET for medical segmentation (90% accuracy), a from-scratch latent diffusion model trained on a Google TPU v5e-8 pod with mobile-budget CPU inference, and VLM fine-tuning for medical visual question answering. My work emphasizes effective fine-tuning (LoRA/QLoRA), robust pipelines, and deployment-ready optimization.