Ghulam Murtaza Portfolio

🔧 Tools and skills:Python, Java, Flask, Heroku, Git, Dockers

- 🔍 Worked with real-world facial images; labeled data using both manual and automated techniques.
- 💡 Built a CNN using TensorFlow for FER 2013 dataset, and leveraged OpenCV's Haarcascade and Adaboost for face detection.
- 📈 Boosted accuracy from 58% to 75% using data augmentation and hyperparameter tuning.
- 🔄 Utilized transfer learning to achieve 89% accuracy on AffectNet dataset (112 GB)
- 🌐 Developed a Flask app to deploy the model; enabled real-time analysis of webcam captured image.
- 🔗 Adopted Git for code versioning and management.
- Deployed an app on Heroku for global accessibility.

🔧 Tools and skills:Python, Java, Flask, Heroku , AWS S3, Git

- 📊 Trained on RAVDESS and SAVEE datasets.
- 🎓 Employed LibROSA for feature extraction from audio.
- 🧠 Fine-tuned CNN model; achieved 70% accuracy.
- 🌐 Real-time predictions on novel voices; 100% accuracy for gender detection.
- 😊😡😢 Detected emotions: calm, happy, sad, angry, fearful, etc.

- 🌐 Developed a web application using Flask
- Managed audio files in AWS S3 via Boto3.
- Processed audio and emotion predictions in the backend.
- Utilized LibROSA for feature extraction; provided real-time results.
- Deployed an app on Heroku for global accessibility.

📊 • Utilized the Facebook Bart model from Hugging Face for summarization.

🎓 • Used Facebook/wav2vec2-large-960h-lv60-self for speech to text transcription.

🌐 • Real-time transcription and summarization of YouTube videos.

🌐 • Web application framework developed in Flask.

💰 • Utilizing the OpenAI DaVinci 003 model for summarization provides high accuracy, but it's quite costly for portfolio inclusion. For the associated code, click here.

📦 • Dockerized the Flask application.

🚀 • Pushed the Docker image to Amazon Elastic Container Registry (ECR).

🛠 • Configured an ECS Task Definition for AWS Fargate.

📡 • Created an ECS Cluster for Fargate.

🚢 • Launched a Service using Fargate, involving configurations including VPC, security groups, and other network settings.

💾 • Allocated 4 vCPUs, 18 GiB soft memory, and 20 GiB hard memory limits.

📡 • Dynamic loading and Automation of app via API Gateway and Lambda function. User visit API which triggers lambda which starts & stops the task.

📡 • Lambda function activates the app for 20 minutes after every API visit using CLoudWatch Events saving $512 by not running 24/7.

⏱ • Added HTML page in Lambda for 7 minute 45 seconds delay during starting of the task, giving update of time left to users for task to initiate.

🚀 • Used AWS Network load balance for consistent IP, ensuring reliable user connection for dynamic fargate ip issues.

🎯 • Load Balancer, listeners, and target groups optimize performance and availability.

📄 • Built a Flask-based web application for document-based question-answering on invoices and receipts.

🔧 • Utilized pdf2image to convert uploaded PDFs into image format for analysis.

🤖 • Integrated the LayoutLM v2 128 million parameters model from Hugging Face's pipeline for detailed extraction from invoices including names, addresses, totals, and subtotals.

💻 • Created a user-friendly web interface, allowing users to upload both image and PDF files for analysis.

⚙️ • Designed custom scripts using jQuery to handle asynchronous file uploads and display analysis results in real-time.

🔄 • Implemented a custom question feature, letting users specify additional queries to the model.

☁️ • Deployed the application on AWS EC2 t2 large instance type and using an Elastic IP and managed version control using Git.

🔧 Using AWS Lambda and CloudWatch Events, I auto-power my EC2 instance from 7 a.m. to 7 p.m EDT MON-FRI, saving ~63% of my costs! ($42 out of total $66) 💰🎉

🔧 • Added the email functionality by using AWS SES(send email service) to send notifications to my email for invocation for trigger events for threshold > 1 , ensuring that the CloudWatch running scripts successfully.

🧠 • Developed a Long-Term Short-Term Memory (LSTM) neural network with an encoder and decoder architecture to translate English sentences to French.

🔧 • Pre-processed the training dataset using tokenization and padding techniques, resulting in an optimized model training process.

📈 • Achieved a 97% accuracy rate on a new dataset of English sentences, showcasing my ability to train and optimize complex neural networks for natural language processing tasks.

📊 • Utilized the F1 score as the primary evaluation metric for model testing, achieving an F1 score of 0.96, Conducted error analysis using confusion matrices and error heatmaps to identify common translation mistakes.