USAMA IMDAD

Professional Experience

• Developed a computer vision pipeline for extracting geological parameters from PDF reports using pdfplumber and OpenCV to create datasets.
• Leveraged self-supervised learning based methods to train foundation models on domain-specific geological image data, improving model performance without requiring large labeled datasets.
• Applied state-of-the-art methods to detect and classify key geological features including defect spacing, inferred strength, color and weathering

• Accomplished a 50% reduction in analysis time, by designing a PyQt software for sugar crystal analysis that utilizes a pre-trained UNET model further trained (Transfer Learning) on a custom dataset for crystal segmentation.
• Developed a person detection system with tracking features for video streams using the TensorFlow JS library and Kalman filter. The system calculates the number of people passing through the camera's view with 80% accuracy.
• Coded a gaze detection algorithm for a hardware device, capable of detecting eye movement in four different directions: left, right, up, and middle. The algorithm achieved high accuracy and was integrated into the hardware.
• Constructed a face recognition system using three models (MTCNN, FaceNet, and SVM) for higher accuracy and faster processing, resulting in a 95%+ accuracy with an inference time of 250ms.
• Accomplished a 10x reduction in false positive rate, as measured by precision, by designing and training a deep learning model to detect fraudulent transactions.
• Fashioned a deep learning model on a project to automate the grading of exams for 1st grade, that achieved 97% accuracy on the test set. As a result, the users were able to grade the exams with less effort and time.

• Drafted front-end and back-end code for a website using React and Node.js, implementing a recommendation system based on collaborative filtering to deliver personalized music recommendations to users.
• Designed a 3D face construction website using Vue JS and Node JS that uses direct volumetric CNN regression on face images to convert them to 3D objects and can return each result in less than 3 minutes.
• Constructed a dashboard using Vue JS and Laravel, cooperating with the team of computer vision developers, for the company's product (a video surveillance system), that managed 3 different purpose computer vision models.

Projects

• Optimized YOLO and ReID models for Jetson Nano and Jetson Orin platforms by implementing FP16 and INT8 quantization, achieving enhanced inference speed and reduced resource consumption without significant accuracy loss.
• Contributed to the BoxMOT repository by fixing a batch size incompatibility in TensorRT-based ReID model inference.
• Developed an n8n workflow integrating LLMs (Gemini) with the WordPress to automate SEO optimization, including dynamic rewriting of titles and meta descriptions to improve search engine visibility and engagement.

Skills

Education

Certifications