Data Scientist | AI Researcher | Machine Learning Engineer
I’m Maimunul Karim Jisan, an aspiring data scientist proficient in developing resource-efficient deep learning models and neural architectures. My expertise spans convolutional neural networks (CNNs), ensemble machine learning algorithms, and advanced statistical methods, including hypothesis testing and regression analysis. Passionate about bridging theoretical knowledge with practical impact, I focus on designing data-driven solutions for sectors like agriculture and healthcare, optimizing model scalability and performance to address real-world challenges such as crop yield prediction and medical diagnostics.
Achieved a Kaggle score of 🏆94.136% accuracy in predicting mental health outcomes using CatBoost. Done real world feature engineering, built a preprocessing pipeline, optimized hyperparameters with Optuna, and compared with multiple model. Evaluated performance with precision (91%), recall (89%), and F1-score (90%). And lastly perform MLOPS in DagsHub for boosting industrial knowledge.
Achieved a Kaggle score of 🏆89.586% accuracy in binary prediction with a Rainfall dataset using Logistic Regression. Created Feature Engineering, built a preprocessing pipeline, optimized hyperparameters with Optuna, and compared with multiple model. Evaluated performance with AUC-ROC (90%) and also used cross validation for model generalisation.
🌾 This project is part of my research journal focused on developing a lightweight model for Rice Disease Classification. We began by collecting data from Mendeley Data 📂 and conducted extensive dataset analysis 🔍. The process started with data augmentation to balance the dataset, followed by building a lightweight model using a Convolutional Neural Network (CNN) 🧠. The model achieved an impressive 99.83%test accuracy ✅ and a 99.37%macro-average across precision, recall, and F1-score 📊. To ensure suitability for low-resource devices, we applied quantization-based model compression ⚙️. After multiple evaluations, we enhanced the model further through Response-Based Knowledge Distillation, which is visualized and explained in the research section below 📘.
Achieved a highest score of 🏆 88.88% accuracy in predicting Bengali Hate Speech outcomes using Logistic Regression 🤖. The dataset was collected from Kaggle 📥 and underwent several preprocessing steps, including emoji removal❌, special character and punctuation removal, stop word filtering 🧹, and tokenization ✂️ and more. We then performed feature engineering ⚙️ and applied upsampling to address class imbalance ⚖️. After vectorization 🧾, we used GridSearchCV 🔍 to fine-tune hyperparameters and built a model pipeline to identify the best-performing algorithm. The final model—Logistic Regression—achieved an exceptional accuracy of 🔥91%. To ensure model transparency and generalization, we also incorporated Explainable AI (XAI)📘 for interpretation and insights.
Achieved a Kaggle rank of 1597/3858 with an accuracy of 95.897% by building a CatBoost model to predict loan defaults. Designed a preprocessing pipeline, optimized the model with GridSearchCV, and utilized both LIME and SHAP for feature importance analysis.
An online non-credit project authorized and offered by HiCounselour, Where developed a Random Forest model to predict cross-sell insurance interest, using SMOTE for class balancing and SHAP for interpretability.
Achieved a Kaggle rank of 69/1558 with an accuracy of 80.827% by building a LightGBM model to predict Spaceship Titanic Data is passenger survived or not. Implemented feature engineering, a preprocessing pipeline, hyperparameters tunning, and model building.
🎯 We achieved a 95.54% test accuracy using the TensorFlow framework. 📥 The dataset was collected from Kaggle, followed by essential image preprocessing steps. 🧠 A custom CNN model was then developed to classify 36 different rice leaf disease classes. After training, the model was evaluated on the test dataset, achieving impressive results: macro precision of 96%, macro recall of 95%, and a macro F1 score of 95%. 🧩 To further enhance deployment efficiency, especially on low-resource devices, we applied a model compression technique called Quantization. 📱 This allowed us to generate a lightweight CNN model that is ideal for real-world, resource-constrained environments.
Developed LiteCNN4Rice, a lightweight deep learning model for rice leaf disease classification, using Response-Based Knowledge Distillation (RKD). The student model achieves 98.15% accuracy with only 16.4K parameters (0.2% of the teacher model size), enabling real-time deployment on IoT devices like Arduino Nano 33 BLE Sense.
First RKD application in agriculture with minimum 11% accuracy boost over standalone models
Optimized for Arduino Nano 33 BLE Sense (0.1MB memory usage)
RLDIS dataset (5,932 images, 4 diseases)
Enables real-time disease diagnosis for smallholder farmers using low-cost devices ($20 Arduino)
Explore the basics of probability distributions, including discrete (e.g., Binomial) and continuous (e.g., Normal) types. Learn how they model data spread, their parameters, and key functions like PMF, PDF, and CDF for data analysis.
Read More
Discover the bell-shaped curve’s role in statistics—symmetry, empirical rule (68-95-99.7), and parameters (mean, standard deviation). See how it’s used in finance, healthcare, and machine learning.
Read More
Learn how Quantile-Quantile (QQ) plots visually compare data against a normal distribution. Interpret deviations (skewness, outliers) and apply this tool to validate assumptions in real-world datasets.
Read More
Unpack the CLT’s power: how sample means form normal distributions regardless of original data shape. Step-by-step examples with the Titanic dataset illustrate its application in statistical inference.
Read More
