Artificial Intelligence Engineer

• Training and Evaluating an Automatic Speech Recognition (ASR) model for Bengali.
• Reduced CER by 5 percent through preprocessing and postprocessing techniques.
• Developed Knowledge Mining solutions using OpenAI GPT models, LangChain, and Azure Cognitive Services.

Machine Learning Engineer

• Trained and developed novel Bangla word-level Optical Character Recognition (OCR) and Bangla Optical Handwriting Recognition (OHR) deep learning models as part of the Enhancement of Bangla Language in ICT through Research & Development (EBLICT) project, under the Ministry of ICT, the Government of Bangladesh.
• Deployed muti-framework models for inference serving on both CPU and GPU servers.
• Converted multi-framework models into ONNX and deployed with NVIDIA Triton Server to reduce complexity and memory optimization.
• Restructured the codebase and converted the API pipelines from Flask to FastAPI to handle more concurrent and parallel requests.
• Worked on Document Denoising to improve the performance of word segmentation and recognition.
• Tailored an Object Detection system to detect the artefacts on the document with few data.
• Developed and deployed and end-to-end solution for Document Classification and Recognition system.

Research Assistant (RA)

• Explored deep learning with synthetic data to mitigate the scarcity of real-world data.
• Trained word recognition module with synthetic data and finetuned with real-world data for higher recognition rate with less real data.
• Studied and developed quantized models to achieve a lower memory footprint and faster inference on the CPU.
• Researched and developed shallow word recognition model using Unsupervised Domain Adaptation.
• Proposed a novel Task Incremental Learning method and published in 26^th International Conference on Pattern Recognition (ICPR).

Software Engineer

• Developed and maintained in-house software.
• Automated multiple processes to reduce user dependencies.
• Reviewed projects plan execution and maintain productivity.

North South University

CGPA: 3.13 on a scale of 4.00 (86%)

A Multifaceted Evaluation of Representation of Graphemes for Practically Effective Bangla OCR

Abstract: Bangla Optical Character Recognition (OCR) poses a unique challenge due to the presence of hundreds of diverse conjunct characters formed by the combination of two or more letters. In this paper, we propose two novel grapheme representation methods that improve the recognition of these conjunct characters and the overall performance of OCR in Bangla. We have utilized the popular Convolutional Recurrent Neural Network architecture and implemented our grapheme representation strategies to design the final labels of the model. Due to the absence of a large-scale Bangla word-level printed dataset, we created a synthetically generated Bangla corpus containing 2 million samples that are representative and sufficiently varied in terms of fonts, domain, and vocabulary size to train our Bangla OCR model. To test the various aspects of our model, we have also created 6 test protocols. Finally, to establish the generalizability of our grapheme representation methods, we have performed training and testing on external handwriting datasets. Experimental results proved the effectiveness of our novel approach. Furthermore, our synthetically generated training dataset and the test protocols are made available to serve as benchmarks for future Bangla OCR research.

Rethinking Task-Incremental Learning Baselines

Abstract: It is common to have continuous streams of new data that need to be introduced in the system in real-world applications. The model needs to learn newly added capabilities (future tasks) while retaining the old knowledge (past tasks). Incremental learning has recently become increasingly appealing for this problem. Task-incremental learning is a kind of incremental learning where task identity of newly included task (a set of classes) remains known during inference. A common goal of task-incremental methods is to design a network that can operate on minimal size, maintaining decent performance. To manage the stability-plasticity dilemma, different methods utilize replay memory of past tasks, specialized hardware, regularization monitoring etc. However, these methods are still less memory efficient in terms of architecture growth or input data costs. In this study, we present a simple yet effective adjustment network (SAN) for task incremental learning that achieves near state-of-the-art performance while using minimal architectural size without using memory instances compared to previous state-of-the-art approaches. We investigate this approach on both 3D point cloud object (ModelNet40) and 2D image (CIFAR10, CIFAR100, MiniImageNet, MNIST, PermutedMNIST, notMNIST, SVHN, and FashionMNIST) recognition tasks and establish a strong baseline result for a fair comparison with existing methods. On both 2D and 3D domains, we also observe that SAN is primarily unaffected by different task orders in a task-incremental setting.

Analysis of Bangla Keyboard Layouts Based on Keystroke Dynamics

Abstract: Keyboards are the primary devices for interaction with computer platforms and have been a central topic in HCI research for decades. The study of their layout design is important in deciding their efficiency, practicality and adoption. Multiple keyboard layouts have been developed for Bangla without any rigorous study for their comparison. In this paper, we take a quantitative data-driven approach to compare their efficiency. Our evaluation strategy is based on the key-pair stroke timing with data collected from standard QWERTY English keyboards. Our experiments conclude that the Bijoy keyboard layout is the most efficient design for Bangla among the four layouts studied. This quantitative approach can lay the groundwork for further study of these layouts based on other criteria.

A satellite collision avoidance system based on General Regression Neural Network

Abstract: Continuous launching of new satellites and increasing numbers of space missions is making space a congested environment. Collision with space debris or other satellites is now a real problem for satellites with the problem more acute in highly trafficked orbits. Thus, mission operators and space agencies are in need of high accuracy collision avoidance systems for spacecrafts and satellites. This paper focuses on tackling the satellite collision problem by implementing a collision avoidance system using neural networks and relevant machine learning techniques. The primary model is based on General Regression Neural Network (GRNN) and the secondary models are based on Artificial Neural Networks (ANN), Random Forest Regression & Support Vector Regression techniques. The dataset used in this paper is collected from the ESA (European Space Agency) which contains the events of risk assessment or in other words, Conjunction Data Messages (CDM). The proposed collision avoidance system predicts the collision risk percentage between target (a satellite of interest) and chaser (space debris or another satellite) objects. The predicted risk enables the target to maneuver accordingly and ultimately avoid collision with the chaser object. The GRNN algorithm uses lazy learning which does not require iterative training and makes predictions based on stored parameters. The training data has been normalized before applying the algorithm as GRNN network is sensitive to high deviation among input features. However, the GRNN model predicts the risk of collision between the target & the chaser object with an MSE (Mean Squared Error) of 11% which means the model predicts the risk of collision with 89% accuracy and this 89% risk can give enough confidence factor to the concerned authority to take necessary evasive maneuvers. This is reliable enough and lower than other models’ MSE to consolidate the fact that the GRNN model is best fit for our dataset.

Soil Analysis and Unconfined Compression Test Study Using Data Mining Techniques

Abstract: In this study, Random Forest Regressor, Linear Regression, Generalized Regression Neural Network (GRNN) and Fully connected Neural Network (FCNN) models are leveraged for predicting unconfined compression coefficient with respect to standard penetration test (N-value), depth and soil type. The study is focused on a particular correlation of undrained shear strength of clay (Cu) with the standard penetration strength. The data used is from 14 no. ward in Mymensingh and Rangamati districts which are situated in Bangladesh. By using this data, the study tries to solidify the correlation of SPT (N-value) with Cu. It also tries to check the goodness of the relationship by comparing it with unconfined compression strength values gained from the unconfined compression test calculated from the field by experts.

Analysis of Soil and Various Geo-technical Properties using Data Mining Techniques

Abstract: In this study, General Regression Neural Network(GRNN), Artificial Neural Network (ANN), Fully Connected Neural Network (FCNN), Support Vector Regression (SVR) and Linear Regression (LR) models have been implemented in order to predict the composition of soil with respect to the Standard Penetration Test (SPT), and soil depth. The primary focus has been on determining a significant correlation between the soil composition with SPT value and depth. Data sets have been used from ward 14, Mymensingh district of Bangladesh and from a construction project along India-Myanmar border. In this study, 8 types of soil, namely, fine sand, silty clay, clayey silt with fine sand, clayey silt, fine sand with silt, silty fine sand, sandy silt, and rubbish has been classified, and the probability of obtaining the soil type classification has been determined.

Forecasting Sea Level Rise Using Machine Learning Techniques

Abstract: During the past few decades, climate change has been posing as a vital game changer for the world stability of natural conditions. The effect can be easily demonstrated via the rise of sea levels on global and local scenarios. Increase of temperature, change in precipitation, melting of glaciers are causing the sea levels to rise in an alarming rate like never before. This particular paper focuses on predicting the sea level of Bangladesh, a third world South Asian regional country using advanced machine learning techniques to produce a potential model for future cautions. The proposed methodology uses climate data of previous 40 years (approx.) from 1977 to 2017 to train our model using different machine learning algorithms like Random Forest (RF), KNN and MLP. In testing phase, KNN algorithm prompted 91.3204% accuracy.

Neural Network from Scratch

A simple fully connected network that can take scalar and vector values and update its gradients using backpropagation, implemented in GoLang from scratch.

Cost Sensitive Learning for Imbalanced Dataset

• A comparative study on different techniques to tackle the problem of imbalanced dataset. We have compared various approaches of handling imbalanced dataset problem and proposed a new loss function that can handle data imbalance problem.
• Wrote a review paper on techniques to deal with imbalanced and long tailed data.

URL Shortener

A web application to generate short URL of any given web address.

Online Classroom

A system for academia where teachers can upload study materials and the system can track the human eye to understand the students reading habits. Can be used as web application and android app.

Smart Parking System

A parking management system where user can enter parking by using their smartphone and the system will allocate an empty parking lot for parking. During checkout, the user will again scan a QR code and the payment will proceed digitally.

Digital Resume Generator

A Java-based digital resume generator which takes information from GUI and generates a web-based digital resume based on given information. Unit testing had been performed in each possible methods.

Live Website Testing Using Selenium

A black box testing is performed on a dummy website phptravels.net using a testing automation tool named Selenium which tested most of the features of the website and generated all test reports using TestNG.

Catch Baby

An android application for parents to track their teenage children's live location. A movement path is generated into the parent’s app from when the parent starts to monitor the activity of the child to the end. Also, the app will send alerts to parents if any child go outside of their permitted zone.

Virtual Library Management System

A library management system to for the admin and user to book, borrow and return and request for books, magazine and journals.

North South University Social Services Club

• Organized “Blood Donation Drive” and successfully collected 573 bags of blood for Bangladesh Red Crescent Society.
• Collected fund and helped 250 flood affected families with dry foods and daily necessity items.
• Arranged winter clothing drive to distribute blankets for 250 underprivileged families in Joypurhaat, Bangladesh.
• Organized “Sociocamp”, an idea generation challenge intending of raising social awareness among the youth.

ACM NSU Student Chapter

• Organized “Technovation”, a 3 days event where over 800 participants, from 22 different universities, competed against each other across six different competitions such as IT Business Challenge, Hackathon, App Development Contest and many more.
• Organized “HackNSU”, where more than 20 teams (each team consisting of 3-4 members) participated from different universities all over the city, with their wide range of thoughts competing against each other over two different problems.