
Omar Sayed Mostafa
Verified Expert in Engineering
AI Engineer and Developer
Cairo, Cairo Governorate, Egypt
Toptal member since August 7, 2023
Omar is a seasoned senior AI engineer with 6+ years of experience in the AI industry, specializing in computer vision, natural language processing, and audio processing. Omar's journey into the world of AI has been fascinating, as he has actively contributed to multiple real-life AI products from inception to deployment. With a combination of hands-on experience and advanced academic training, Omar is well-equipped to tackle any AI challenge.
Portfolio
Experience
- Python - 5 years
- Computer Vision - 5 years
- Machine Learning - 5 years
- PyTorch - 4 years
- Video Analysis - 3 years
- Object Detection - 3 years
- Natural Language Processing (NLP) - 2 years
- Audio Analysis - 1 year
Availability
Preferred Environment
Linux, Git, PyTorch, Python, Docker, Natural Language Processing (NLP), Computer Vision, Runtime Optimization, Neural Networks
The most amazing...
...project I've led was at precision.ai. It involved diving into the architectural optimization of complex AI models, tailoring them for Jetson boards.
Work Experience
Machine Learning Engineer
Kog SAS
- Developed a high-performance inference framework for large language models (LLMs) leveraging AMD GPUs, utilizing HIP and C++.
- Contributed to the development of an AI-driven game creation platform enabling users to input game scenarios, which the AI then transforms into game assistance, visual game graphs, game logic, and executable code.
- Focused on prompt engineering and fine-tuning large language models to align with specific product requirements.
Senior AI Engineer
Rosalyn
- Collaborated closely with a diverse, cross-functional, multinational team to translate intricate product requirements into advanced AI/ML models and architect innovative solutions.
- Worked on developing retrieval-augmented generation (RAG) applications leveraging large language models (LLMs) through various frameworks such as Hugging Face, LangChain, and OpenAI APIs.
- Was continuously involved in fine-tuning prompts (prompt-tuning) tailored for different large language models (LLMs) to adapt to the evolving downstream tasks of the product seamlessly.
- Developed video analysis modules to detect and analyze different video activities in real-time streams, including object detection, object tracking, pose detection, gaze detection, segmentation, and motion detection.
- Built models for voice activity detection and speech recognition (detecting and converting speech to text) within audio streams.
- Maintained a multiprocessing multithreaded environment, enabling parallel execution of diverse AI models to significantly improve performance and reduce processing time.
- Enhanced and optimized current working AI systems with end-to-end responsibility, including design, code implementation, training, testing, and deployment.
- Implemented API endpoints to enable AI model inference, fostering real-time predictions and seamless integration.
- Containerized models using Docker for streamlined deployment.
Senior Computer Vision Engineer
Silver Bullet Services Group
- Collaborated remotely with a multinational team on a video-text retrieval task via Contrastive Language-image Pre-training learning and audio classification.
- Worked on auto-greedy label generation for large datasets and used active learning to optimize automatic labels, resulting in faster annotation and improved model performance.
- Fine-tuned large complex models, including a language encoder and an image encoder, using the customized dataset with generated labels, optimizing model performance for specific tasks.
Senior Machine Learning Engineer
Botit
- Improved existing NER models for Arabic and English languages by studying the behavior of the models and spotting model weaknesses and lack of information representation in the training data.
- Generated a balanced synthetic robust representation of entities and intents across different domains to represent ambiguity in the language and cover the corner cases the model failed to understand.
- Updated a labeling schema for the training data to give the model more information during training enabling it to understand the context of the input text and differentiate between different domains of ambiguity.
- Implemented training codebase using the Rasa platform for training transformer-based AI models.
Senior Machine Learning Engineer
Precision.ai
- Researched and applied SOTA approaches to detecting and spraying weeds across different crops in different fields, leveraging supervised, self-supervised, and intermediate-supervised learning while working alongside a multinational team.
- Implemented and trained autoencoder models for image semantic segmentation, in addition to using transformer-based models to apply knowledge distillation learning.
- Worked on a model runtime optimization project that included model architecture tweaks and quantization-aware training to increase target model throughput and decrease latency while maintaining accuracy.
- Led a small team responsible for optimizing a model–architecture-wise–to satisfy runtime requirements on high-end devices such as NVIDIA Jetson boards while maintaining accuracy.
- Utilized ONNX, TensorRT, and TorchScripts to provide enhanced and efficient model engines.
- Developed an end-to-end pipeline model, including training, evaluation, and deployment of end devices.
- Implemented parallel training procedures using PyTorch (Distributed Data Parallel) to speed up the training process across multiple GPUs and multiple nodes, which saved more than 25% on Amazon EC2 instances training costs.
- Provided semi-automated methods for semi-automated annotation of large crop datasets with minimal human interaction, leveraging active learning and other classic approaches to speed up the annotation process and save annotating costs from scratch.
Computer Vision Engineer
Hansa Robotics
- Contributed to the optimization and quantization of object detection models to fit the performance of edge devices.
- Trained and deployed real-time object detection models (YOLO) on custom datasets using PyTorch.
- Collected and annotated custom datasets from real product sources and generated further synthetic data to increase and balance the representation of datasets used in training.
- Deployed object detection models (YOLO) on edge boards (Jetson boards) using TensorRT and DeepStream pipeline.
- Converted PyTorch models using ONNX to TensorRT engine.
Experience
PyTorch Segmentation for Upper and Lower Jaws
https://github.com/OmarSayedMostafa/Face-analysisDeep Learning Colorization for Visual Media
https://github.com/OmarSayedMostafa/Deep-learning-Colorization-for-visual-mediaEducation
Master's Degree in Computational Linguistics
Helwan University - Cairo, Egypt
Bachelor's Degree in Computer Science
Ain Shams University - Cairo, Egypt
Skills
Libraries/APIs
PyTorch, OpenCV, TensorFlow, Natural Language Toolkit (NLTK), SpaCy, Rasa NLU
Tools
Git, Open Neural Network Exchange (ONNX), You Only Look Once (YOLO), ChatGPT, TensorBoard, Whisper, Rasa.ai
Languages
Python, Python 3, C++
Paradigms
Parallel Programming, Distributed Computing
Platforms
Linux, Amazon EC2, Docker, NVIDIA CUDA, Apache Kafka, AMD
Storage
Amazon S3 (AWS S3)
Frameworks
GStreamer
Other
Natural Language Processing (NLP), Deep Learning, Computer Vision, Image Processing, Object Detection, Semantic Segmentation, Artificial Intelligence (AI), Machine Learning, Supervised Machine Learning, Large Language Models (LLMs), Conda, Neural Networks, Videos, Labeling, Data Preprocessing, Active Learning, Audio, NVIDIA TensorRT, Natural Language Understanding (NLU), Video Analysis, Audio Analysis, Image Recognition, Generative Pre-trained Transformers (GPT), Visualization, APIs, Speech Recognition, Speech to Text, LangChain, Image Search, Recurrent Neural Networks (RNNs), OpenAI, Generative Pre-trained Transformer 3 (GPT-3), Retrieval-augmented Generation (RAG), Data Engineering, Scientific Computing, Runtime Optimization, Prompt Engineering, LoRa, PEFT, ChatGPT Prompts, Art, Creativity, Models, HIP/CUDA
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
Share your needs
Choose your talent
Start your risk-free talent trial
Top talent is in high demand.
Start hiring