I am a machine learning engineer who is passionate about using AI to make the world a better place. As a data scientist, I thrive on using statistics to gain insights into the world and make meaningful contributions through the development of AI tools that simplify life.
Connect with me on LinkedIn or reach me at dahshury@gmail.com
A gym back machine detector (7 machines) using RT-DETRv2 and YOLOv8 with images from various online datasets. This can be used in real-world scenarios to identify gym machines even when humans are using them. The solution is deployed on Microsoft Azure's Web Apps for containers, with a front-end built using Streamlit.
2. WinSTT
An application for desktop Speech-to-Text (STT) using ONNX Whisper-Turbo, a lightweight STT model based on OpenAI’s Whisper. The application offers fast transcription and supports over 99 languages without an internet connection. The UI, developed with PyQT6, was compiled into a ".exe" file for easy use, including customizable hotkeys and fast transcription. Ideal for writers or daily use.
Estimating the positions of 25 keypoints on the human body by training an Hourglass network (from scratch), YOLOv8 Pose, and Keypoint RCNN. The model achieves 68% mAP@50 on the validation set, helping to solve problems such as pose tracking and gesture recognition.
A gym back machine detector (7 machines) using RT-DETRv2 and YOLOv8 with images from various online datasets. This can be used in real-world scenarios to identify gym machines even when humans are using them. The solution is deployed on Microsoft Azure's Web Apps for containers, with a front-end built using Streamlit.
A marble quality classifier implemented using:
- From-scratch implementation of ResNet34.
- Pretrained ResNet50.
- Huggingface ViT transformer.
A face recognition system using a MySQL database to store and retrieve face embeddings, providing a robust database-backed solution for secure identification and authentication.
A classification task predicting whether a customer will subscribe to a term deposit based on 12 features.
A regression task predicting the prices of Australian vehicles based on 12 features using Exploratory Data Analysis (EDA), data cleaning, and feature extraction. The solution is deployed on GCP App Engine with a Streamlit front end.
A vehicle segmentation project using UNET (from scratch), achieving 0.99-pixel accuracy. The project focuses on separating the vehicle from the background, which is useful for applications in autonomous driving and image analysis.