Muaz is available for hire
Hire MuazMuaz Maqbool
Verified Expert in Engineering
Artificial Intelligence Developer
Location
Lahore, Punjab, Pakistan
Toptal Member Since
June 8, 2023
Muaz has over five years of product development experience, having co-founded two AI startups that raised over $200,000. As the former CTO of a service-based company, he executed over 50 projects. Muaz has contributed to the industry's academic discourse with 11 published papers as an industrial research supervisor. He applies his in-depth knowledge and expertise as a remote consultant, establishing over 20 strategic engagements with US-based enterprises.
Portfolio
Self-employed
Artificial Intelligence (AI), Computer Vision, Machine Learning...
Freede Solutions Inc.
Python, Artificial Intelligence (AI), Machine Learning...
Hao Ting Yen (Scoop AI)
Python, Machine Learning, APIs, OpenAI Gym, Natural Language Processing (NLP)...
Experience
Availability
Part-time
Preferred Environment
Artificial Intelligence (AI), Visual Studio Code (VS Code), Computer Vision, Natural Language Processing (NLP), Data Science
The most amazing...
...thing I've developed is SportsEye—an advanced real-time stats computation engine for soccer broadcast videos—using a complex 8-model deep learning pipeline.
Work Experience
Computer Vision and NLP Expert | AI Consultant and Contractor
2022 - PRESENT
Self-employed
- Served as a consultant and contractor. Brought in technical expertise and managerial skills to assist 20+ US companies in seamlessly integrating AI into their products.
- Built and maintained over 25 long-term engagements as an expert in the Expert Vetted Batch (Top 1%) category with a remarkable 99% job success rate.
- Leveraged OpenAI LLMs to deliver solutions, including automated property description generation, chatbot development, and embedding-based engines. Created models for medical insurance billing, contributing to the success of multiple businesses.
Technologies: Artificial Intelligence (AI), Computer Vision, Machine Learning, Natural Language Processing (NLP), Data Science, GitHub, Machine Learning Operations (MLOps), AWS CLI, Google Cloud Platform (GCP), Language Models, FastAPI, Flask, Generative Pre-trained Transformer 3 (GPT-3), OpenAI GPT-4 API, OpenAI GPT-3 API, OpenCV, PyTorch, ChatGPT, Generative Pre-trained Transformers (GPT), Pandas, Jupyter, Programming, Integration, Natural Language Toolkit (NLTK), SpaCy, LangChain, GPT, Amazon Web Services (AWS), Image Recognition, Visualization, Image Processing, Web Scraping, Back-end Development, Google Vision API, Image Search, Labeling, Amazon Rekognition, Fashion, PostgreSQL, Recurrent Neural Networks (RNNs), Neural Networks, Text to Image, Convolutional Neural Networks (CNN), You Only Look Once (YOLO), Open Neural Network Exchange (ONNX), Regression Modeling, Quantitative Analysis, Forecasting, FFmpeg, Research, Communication, Generative Adversarial Networks (GANs), OpenAI, Custom Models, Generative Artificial Intelligence (GenAI), Architecture, Industrial Internet of Things (IIoT), Time Series Analysis, Transformer Models, Time Series, Generative AI, Retrieval-augmented Generation (RAG), Signal Processing, Data Extraction, PDF, Databases, Cloud, Software Architecture, Algorithms, Models, Data, AI Integration, Prompt Engineering
AI Expert
2023 - 2024
Freede Solutions Inc.
- Created a state-based chatbot for debt collection, navigating through different stages during the debt collection process, including guardrails to avoid unexpected behavior.
- Fine-tuned various LLMs for different generative and multilabel classification tasks for various learning techniques such as fine-tuning, few-shot learning, and zero-shot learning.
- Evaluated more than 15 LLMs of different sizes (7B - 180B parameters) for fixed debt collection scenarios.
- Generated conversational, synthetic data for various tasks.
Technologies: Python, Artificial Intelligence (AI), Machine Learning, Generative Pre-trained Transformers (GPT), OpenAI, Chatbots, Chatbot Conversation Design, Natural Language Processing (NLP), AIOps, RunPod, Serverless, APIs, Deployment, Training, Modeling, AI Chatbots, Time Series Analysis, Transformer Models, Time Series, Generative AI, PDF, Databases, Cloud, Software Architecture, Models, Data, AI Integration, LangChain, Prompt Engineering
NLP Engineer
2023 - 2024
Hao Ting Yen (Scoop AI)
- Developed an automated pipeline to scrap news from various internet sources and categorized them among multiple categories. Deployed in on AWS Cloud Instance.
- Generated podcast transcripts using Open AI LLMs with scrapped articles from various sources.
- Implemented text to speech (TTS) feature to generate podcast audio clips using the transcripts generated by OpenAI LLMs.
Technologies: Python, Machine Learning, APIs, OpenAI Gym, Natural Language Processing (NLP), Natural Language Toolkit (NLTK), SpaCy, Artificial Intelligence (AI), Language Models, OpenAI, Custom Models, Generative AI, PDF, Databases, SQL, Cloud, Software Architecture, Models, Data, AI Integration, LangChain, Prompt Engineering
Computer Vision Researcher
2023 - 2023
Pixelcut Inc.
- Implemented TryOnDiffusion, the latest paper on virtual try-on from scratch.
- Implemented cross- and self-attention along with a cascaded diffusion pipeline.
- Conducted research on top virtual try-on models and did comparisons for available open-source projects, including CP-VTON, CP-VTON+, and HR-VITON.
Technologies: Computer Vision, TensorFlow, PyTorch, Machine Learning, Artificial Intelligence (AI), Generative Adversarial Networks (GANs), Custom Models, OpenCV, Databases, Cloud, Models, Data
Computer Vision Contractor
2021 - 2022
Agot
- Partnered with the CTO to improve accuracy for detection and tracking methods. Integrated the new pipeline into their system.
- Improved multi-object tracking accuracy for kitchen menu items by 3%. Focused on reducing ID switches per ground truths and tried different weighted cost functions and heuristic-based improvements.
- Increased multi-object detection mAP by 30% by modifying YOLOv5 classification loss to support multi-label object detection.
Technologies: Computer Vision, PyTorch, Machine Learning, Jira, Slack, GitHub, Programming, Amazon Web Services (AWS), Image Recognition, Image Processing, Google Vision API, Image Search, Labeling, PostgreSQL, Recurrent Neural Networks (RNNs), Neural Networks, Convolutional Neural Networks (CNN), You Only Look Once (YOLO), Quantitative Analysis, FFmpeg, Communication, Generative Adversarial Networks (GANs), Custom Models, OpenCV, Databases, Cloud, Models, Data
Product Development Lead | CPO
2020 - 2022
Adlytic
- Developed audience analytics through CCTV footage for a retail store. Built the pipeline using machine learning and deep learning algorithms with PyTorch, Keras, and TensorFlow in Python.
- Deployed a scalable solution using Docker to process multiple cameras. The product included features such as footfall counting, dwell time, heatmaps generation, and age, gender, and emotion classification.
- Implemented the product's machine learning and deep learning models in PyTorch and TensorFlow and later optimized using the C++-enabled framework, TensorRT.
Technologies: Jira, GitHub, Slack, Management, Computer Vision, Deployment, Machine Learning Operations (MLOps), NVIDIA CUDA, Jetson TX2, Artificial Intelligence (AI), Python, PyTorch, TensorFlow, Deep Learning, Machine Learning, Algorithms, Keras, Docker, Jupyter, Programming, Integration, Amazon Web Services (AWS), Image Recognition, Back-end Development, Neural Networks, Convolutional Neural Networks (CNN), You Only Look Once (YOLO), Regression Modeling, Quantitative Analysis, Leadership, Communication, Custom Models, OpenCV, Industrial Internet of Things (IIoT), Databases, SQL, Software Architecture, Models, Data
AI Lead | CTO
2019 - 2022
Omno Ai
- Spearheaded the development and launch of OMNO AI's flagship products, Adlytic, Trafflytic, and SportsEye, leveraging AI to revolutionize person, traffic, and sports data analytics for videos and live streams.
- Engineered a cutting-edge, scalable traffic analytics solution deployed for the Indonesian and Turkish governments. Features included real-time traffic density analysis, classified directional counting, and precise lane counting capabilities.
- Increased the AdMob engine's conversion by 200% by creating a scalable hybrid AdMob recommendation engine with collaborative and content-based filtering. Processed one million inferences daily, with a 24-hour online learning cycle.
- Delivered 50+ cutting-edge AI projects, expertly leading a team of eight data scientists and DevOps engineers from pre-sales to deployment. Conducted thorough code reviews and provided invaluable management expertise.
Technologies: AIOps, Artificial Intelligence (AI), Computer Vision, Data Science, Google Cloud Platform (GCP), Machine Learning, GitHub, AWS CLI, Deep Learning, APIs, Dashboards, Management, Jira, Slack, Deployment, Machine Learning Operations (MLOps), Natural Language Processing (NLP), ChatGPT, Generative Pre-trained Transformers (GPT), Pandas, Financial Modeling, Jupyter, Programming, Integration, Natural Language Toolkit (NLTK), SpaCy, Amazon Web Services (AWS), Image Recognition, Image Processing, Web Scraping, Image Search, Labeling, Amazon Rekognition, PostgreSQL, Recurrent Neural Networks (RNNs), Neural Networks, Convolutional Neural Networks (CNN), You Only Look Once (YOLO), Open Neural Network Exchange (ONNX), Regression Modeling, Quantitative Analysis, Forecasting, FFmpeg, Leadership, Communication, Generative Adversarial Networks (GANs), Custom Models, OpenCV, Time Series Analysis, Transformer Models, Time Series, Databases, Software Architecture, Algorithms, Models, Data
Experience
SportsEye
https://www.youtube.com/watch?v=09Hik9FpzFM&t=53sAn advanced computer vision-enabled deep learning product designed for soccer analysis. It processes real-time soccer broadcast matches, providing comprehensive statistical insights with 24 locality and possession metrics. The system is built upon the following components:
• Goal or attempts event detection using CALF and 3D ResNets.
• Field line segmentation using Pix2Pix-based GANs.
• Player and ball detection using the YOLO SSD family.
• Player re-identification and tracking using visual features and Kalman filter-based tracking methods.
• Camera view classification using EfficientNet models.
• Field key points localization using FLANN-based retrieval on Siamese-based features for field lines and classical holography techniques to estimate camera poses, enabling 3D perspective visualization and enhancing virtual game simulations.
• Timer localization and OCR-based recognition using EAST and EasyOCR-based methods.
With its extensive functionalities and cutting-edge technologies, SportsEye revolutionizes soccer analysis by providing real-time insights and precise statistical data.
• Goal or attempts event detection using CALF and 3D ResNets.
• Field line segmentation using Pix2Pix-based GANs.
• Player and ball detection using the YOLO SSD family.
• Player re-identification and tracking using visual features and Kalman filter-based tracking methods.
• Camera view classification using EfficientNet models.
• Field key points localization using FLANN-based retrieval on Siamese-based features for field lines and classical holography techniques to estimate camera poses, enabling 3D perspective visualization and enhancing virtual game simulations.
• Timer localization and OCR-based recognition using EAST and EasyOCR-based methods.
With its extensive functionalities and cutting-edge technologies, SportsEye revolutionizes soccer analysis by providing real-time insights and precise statistical data.
Adlytic
https://adlytic.ai/An AI-powered retail analytics technology that analyzes the store's audience analytics through CCTV footage using detection and tracking models. I built its computer vision pipeline using machine learning and deep learning algorithms in Pything with PyTorch, Keras, and TensorFlow. The solution is deployed using Docker, is scalable to process multiple cameras, and involves the following features:
• Footfall counting
• Dwell time
• Heatmaps generation
• Age, gender, and emotion classification
• Area-wise conversion
• Staff or customer classification
• Intrusion detection
This product's machine learning and deep learning models were implemented in PyTorch and TensorFlow, which were later optimized using the C++-enabled framework, TensorRT.
• Footfall counting
• Dwell time
• Heatmaps generation
• Age, gender, and emotion classification
• Area-wise conversion
• Staff or customer classification
• Intrusion detection
This product's machine learning and deep learning models were implemented in PyTorch and TensorFlow, which were later optimized using the C++-enabled framework, TensorRT.
Smart Gandola
Smart Gandola incorporates an audience detector based on cutting-edge computer vision algorithms. This set up seamlessly captures frames and transmits them to the Cloud-based pipeline deployed on AWS EC2. On the cloud instance, I deployed a sophisticated pipeline that integrates state-of-the-art models, including face detection, mask detection, frontal face classifier, age classification, gender classification, and emotion classification.
Using these powerful algorithms, Smart Gandola generates comprehensive statistics such as total traffic count, dwell time, and gaze time. These metrics provide invaluable insights into consumer behavior, enabling businesses to make data-driven decisions and optimize their advertising strategies.
To enhance personalization, I implemented advanced age and gender profiling techniques within Smart Gandola. By developing a user-friendly dashboard, businesses can now have precise control over advertisements, allowing them to assign specific ads to predefined age and gender groups. This level of personalization ensures that the right message reaches the right audience, maximizing the impact and return on investment for advertising campaigns.
Using these powerful algorithms, Smart Gandola generates comprehensive statistics such as total traffic count, dwell time, and gaze time. These metrics provide invaluable insights into consumer behavior, enabling businesses to make data-driven decisions and optimize their advertising strategies.
To enhance personalization, I implemented advanced age and gender profiling techniques within Smart Gandola. By developing a user-friendly dashboard, businesses can now have precise control over advertisements, allowing them to assign specific ads to predefined age and gender groups. This level of personalization ensures that the right message reaches the right audience, maximizing the impact and return on investment for advertising campaigns.
Ad-mob Recommendation Engine
By combining collaborative and content-based filtering techniques with the knowledge base, this powerful engine ensures the delivery of highly relevant and personalized recommendations.
Implemented using PySpark on AWS EMR, the engine updates the collaborative and content-based filters every 24 hours to provide the best recommendations daily. The FAST Inference API, hosted on EC2, scales efficiently using a Kubernetes cluster with multiple pods to handle over 8 million inferences per day while meeting the efficiency criteria of 300ms per request.
To support efficient inference, a Redis instance is used to cache results. The deployment strategy follows the green-blue approach to minimize downtime and ensure uninterrupted service during updates after each online learning cycle.
An online learning mechanism was set in place using the green-blue deployment strategy. The system was able to improve the conversion of the Ad Mob engine by 200%
Implemented using PySpark on AWS EMR, the engine updates the collaborative and content-based filters every 24 hours to provide the best recommendations daily. The FAST Inference API, hosted on EC2, scales efficiently using a Kubernetes cluster with multiple pods to handle over 8 million inferences per day while meeting the efficiency criteria of 300ms per request.
To support efficient inference, a Redis instance is used to cache results. The deployment strategy follows the green-blue approach to minimize downtime and ensure uninterrupted service during updates after each online learning cycle.
An online learning mechanism was set in place using the green-blue deployment strategy. The system was able to improve the conversion of the Ad Mob engine by 200%
Medical Insurance Billing Advisor GPT-3-based Chatbot
This project involved the fine-tuning of the powerful Divinchi-03 model on the specific guidelines set forth by the American Medical Association (AMA) for the year 2023. Leveraging the OpenAI API, I embarked on a meticulous process that involved the creation of a comprehensive question-and-answer (Q/A) dataset tailored to the AMA's guidelines.
By fine-tuning the Divinchi-03 model using this carefully curated dataset, I ensured that the model possessed a deep understanding of the latest medical practices and recommendations outlined by the AMA. This state-of-the-art model became an invaluable resource for healthcare professionals, offering accurate and up-to-date answers to a wide range of medical queries.
To provide seamless access and integration for end-users, I integrated the fine-tuned model into the client dashboard of True CRM, a prominent medical insurance billing company based in the United States. This integration enabled True CRM's users to directly access the refined model's capabilities, enhancing their efficiency and accuracy in dealing with medical billing and related inquiries.
By fine-tuning the Divinchi-03 model using this carefully curated dataset, I ensured that the model possessed a deep understanding of the latest medical practices and recommendations outlined by the AMA. This state-of-the-art model became an invaluable resource for healthcare professionals, offering accurate and up-to-date answers to a wide range of medical queries.
To provide seamless access and integration for end-users, I integrated the fine-tuned model into the client dashboard of True CRM, a prominent medical insurance billing company based in the United States. This integration enabled True CRM's users to directly access the refined model's capabilities, enhancing their efficiency and accuracy in dealing with medical billing and related inquiries.
Automated Podcast Generation Using OpenAI LLMs
Automated a pipeline with the following features:
• Scrap news data from various Internet sources;
• Categorise data into multiple verticals (Healthcare, Politics, Sports, Tech);
• Summarize news content using OpenAI LLMs;
• Create podcasts using OpenAI LLMs;
• Create audio podcast clips using text-to-speech (TTS).
• Scrap news data from various Internet sources;
• Categorise data into multiple verticals (Healthcare, Politics, Sports, Tech);
• Summarize news content using OpenAI LLMs;
• Create podcasts using OpenAI LLMs;
• Create audio podcast clips using text-to-speech (TTS).
Debt Collection Assistant
https://www.freede.co/In this project, we developed a state-of-the-art chatbot specifically designed for the debt collection industry. This intelligent system is engineered to navigate through the various stages of the debt collection process with high efficiency and sensitivity. The chatbot incorporates advanced guardrails to prevent any unforeseen behaviors, ensuring a smooth and respectful interaction with the clients.
1. State-based navigation
2. Custom-tailored language models
3. Extensive model evaluation
4. Synthetic conversational data generation
The development of this chatbot marks a significant leap forward in automating the debt collection process. By leveraging cutting-edge AI and language models, we have created a system that not only improves the efficiency of debt recovery but also ensures a respectful and constructive engagement with clients. This project represents a blend of technological innovation and practical application, setting a new standard in the field of AI-driven debt collection.
1. State-based navigation
2. Custom-tailored language models
3. Extensive model evaluation
4. Synthetic conversational data generation
The development of this chatbot marks a significant leap forward in automating the debt collection process. By leveraging cutting-edge AI and language models, we have created a system that not only improves the efficiency of debt recovery but also ensures a respectful and constructive engagement with clients. This project represents a blend of technological innovation and practical application, setting a new standard in the field of AI-driven debt collection.
Education
2023 - 2023
Master's Degree in Artificial Intelligence
Georgia Institute of Technology - Atlanta, Georgia, USA
2015 - 2019
Bachelor's Degree in Computer Science
National University of Computer and Emerging Sciences - Lahore, Punjab, Pakistan
Skills
Libraries/APIs
PyTorch, TensorFlow, OpenCV, Pandas, Natural Language Toolkit (NLTK), SpaCy, Google Vision API, Amazon Rekognition, FFmpeg, Keras, PySpark
Tools
GitHub, Slack, Jupyter, You Only Look Once (YOLO), AWS CLI, Jira, ChatGPT, Jetson TX2, Git, Amazon Elastic MapReduce (EMR), AWS Deployment, OpenAI Gym
Languages
Python, SQL, Python 3
Paradigms
Data Science, Management, REST
Platforms
Visual Studio Code (VS Code), Amazon Web Services (AWS), Android, Google Cloud Platform (GCP), NVIDIA CUDA, Docker, Amazon EC2, Raspberry Pi, Kubernetes
Storage
PostgreSQL, Databases
Frameworks
Flask
Other
Artificial Intelligence (AI), Software, Computer Vision, Machine Learning, Natural Language Processing (NLP), Deep Learning, Video Analysis, Generative Pre-trained Transformers (GPT), Programming, AI Programming, GPT, Image Recognition, Visualization, Image Processing, Back-end Development, Image Search, Labeling, Recurrent Neural Networks (RNNs), Neural Networks, Convolutional Neural Networks (CNN), BERT, Open Neural Network Exchange (ONNX), Regression Modeling, Quantitative Analysis, Forecasting, Research, Communication, Generative Adversarial Networks (GANs), OpenAI, Custom Models, Generative Artificial Intelligence (GenAI), Large Language Models (LLMs), Architecture, AI Chatbots, Time Series Analysis, Transformer Models, Time Series, Generative AI, Retrieval-augmented Generation (RAG), Data Extraction, PDF, Software Architecture, Models, Data, AI Integration, AIOps, Algorithms, Generative Pre-trained Transformer 3 (GPT-3), OpenAI GPT-3 API, OpenAI GPT-4 API, Language Models, FastAPI, Financial Modeling, Integration, LangChain, Web Scraping, Fashion, Text to Image, Leadership, Industrial Internet of Things (IIoT), Cloud, Prompt Engineering, Machine Learning Operations (MLOps), APIs, Dashboards, Deployment, GPU Computing, Recommendation Systems, Big Data, Autoscaling Groups, Chatbots, OCR, Videos, Chatbot Conversation Design, RunPod, Serverless, Training, Modeling, Signal Processing
Collaboration That Works
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
1
Share your needs
Discuss your requirements and refine your scope in a call with a Toptal domain expert.
2
Choose your talent
Get a short list of expertly matched talent within 24 hours to review, interview, and choose from.
3
Start your risk-free talent trial
Work with your chosen talent on a trial basis for up to two weeks. Pay only if you decide to hire them.
Top talent is in high demand.
Start hiring