Ahmed Fakhry

Leveraging large language models for vision applications such as image captioning and visual question answering for surveillance applications. Developed a highly effective video anomaly detection system utilizing few-shot and meta-learning techniques, achieving an AUC of 86.0% and demonstrating robust generalization to previously unseen anomalous actions. Collaborate on multiple computer vision projects with Korean research institutes and private companies to develop innovative technologies and optimize existing surveillance systems. Analyze large datasets using statistical methods to derive insights, identify patterns, and validate models with a focus on raising efficiency, accuracy, and scalability.

Ensured high-quality outputs and results by working closely with approx. 10 AI trainers and 5 LLM prompt engineers to improve the model resilience against hallucinations and factual inaccuracies. Contributed to a data science team focused on optimizing large language models like ChatGPT 3.5 by applying Reinforcement Learning from Human Feedback (RLHF) to human prompts and review data, enhancing the quality of user interactions. Partnered with several large AI companies, leveraging collaborative efforts to integrate advanced technologies and drive innovation within the team's projects.

Created a deep learning model specifically designed to digitize handwritten documents in English and Arabic by leading approx. 5 Data Scientists, enhancing text recognition capabilities. Refined the model's ability to recognize and digitize text accurately by training the model on thousands of handwritten documents, achieving a remarkable CER of 0.24 for English and 0.33 for Arabic. Built and supervised an in-house Annotation team to annotate handwritten text in both languages and provide high-quality training data for supervised learning. Surpassed industry benchmarks and outperformed the Google OCR tool by fine-tuning the models performance in digitizing handwritten documents. Drove B2B engagements in the medical industry across Egypt and Saudi Arabia while engaging with key stakeholders from insurance and hospital companies to foster strategic partnerships and expand business opportunities.

Managed 4 Machine Learning Engineers to formulate an AI-based method for precise prediction of COVID-19 infection rates with a ROC-AUC of 77.1%, facilitating data-driven decision-making. Rendered strategic direction for the creation of a state-of-the-art deep learning model for detecting COVID-19 on the COUGHVID dataset while presenting the results through a research paper at the Interspeech Conference 2021. Acquired clean and standardized clinical datasets to progress project research and reach substantial conclusions by negotiating agreements with hospitals. Enabled optimal integration and functionality by overseeing all aspects of deploying machine learning models in production, including making decisions regarding performance, fine-tuning based on clinical datasets, and testing APIs.

Participated in a research project aimed at establishing an advanced deep learning model for detecting gene-gene interactions from gene expression data, contributing to advancements in genetic research methodologies. Secured an exceptional AUROC score of 0.834 on the combined BioGRID and DREAM5 dataset, enabling the model to achieve a current ranking as the best in performance.

Created a multi-branch deep learning network, integrating heterogeneous features from patients, such as human cough sounds and metadata (age/gender), to detect respiratory diseases, including common cold and COVID-19. Fulfilled real-time requirements by adapting machine learning and neural network algorithms and architectures for optimal performance in dynamic environments.