Mahedi Hasan Rasel

This project was designed for a Japanese automotive parts manufacturer to ensure high standards of quality and accuracy in their packaging. The goal was to automate the process of detecting and measuring specific holes on packaging boxes to verify if they adhere to strict dimensional requirements.

The system employs advanced computer vision algorithms to analyze each box cover. It begins by detecting the precise location of each hole, even under variations in lighting or surface texture. Once a hole is identified, the algorithm calculates its diameter and other spatial coordinates, as well as its distance to defined reference points.

By automating this quality control step, the system reduces the need for labor-intensive manual inspections, speeds up production, and enhances accuracy. This level of automation is particularly valuable for the automotive industry, where precision is critical.

This project addresses a critical challenge in industrial tool manufacturing: ensuring every tool is present and accounted for in production boxes. Using advanced image processing and a custom-trained CNN model, the software accurately detects avariety of tools, including screwdrivers, hammers, and other specialized equipment.

This project leverages the power of LLama3.1 to automate the creation of personalized cold emails. The system analyzes job postings and company information to generate contextually relevant and engaging email content.

This system combines facial recognition and OCR technologies to automate passport number tracking and verification using VGGFace for face recognition and MTCNN for face detection.

Developed and implemented a Deep Convolutional Generative Adversarial Network (DCGAN) using TensorFlow and Keras to generate synthetic images based on the MNIST dataset, demonstrating image quality improvement over training epochs.

Utilizes Pegasus model to automate text summarization. Achieved ROUGE-1 at 0.613, ROUGE-2 at 0.373, ROUGE-L at 0.557, and ROUGE-Lsum at 0.556.

BERT-based model for multi-label classification of user feedback. Achieved F1-scores of 0.84 (micro), 0.81 (macro), and 0.83 (weighted) across 415 samples.

Revenue forecasting using SARIMA and LSTM models. SARIMA RMSE: 529.56, LSTM RMSE: 323.89.

CNN-based system for detecting and classifying signs related to accessibility and disability services.

Utilizes SHAP and LIME to visualize and interpret model predictions across tabular, text, and image data, building trust in machine learning models.

MLflow implementation for tracking and managing machine learning experiments focused on predicting waiting times in service operations.

Explores Named Entity Recognition using state-of-the-art language models including MISTRAL, BERT, and FLAN T5, comparing performance across different pre-trained models.

Multimodal project combining CNN and RNN architectures to automatically generate descriptive captions for images across various scenarios.