MSA-TransUNet++: GI Tract Segmentation in MRI
A hybrid CNN and transformer for segmenting the stomach, small bowel and large bowel in abdominal MRI. Multi-scale attention modules capture channel and spatial dependencies at 3×3, 5×5 and 7×7 receptive fields, and an 8-layer Swin Transformer bottleneck models long-range context. Evaluated on the UW-Madison GI Tract benchmark, with a combined loss for heavy class imbalance.
Segmentation
PyTorch
Swin Transformer
MRI
Inception Swin-UNet for Micro-CT Segmentation
Semantic segmentation of micro-CT rock volumes into void, grain and calcite. An Inception-style Swin-UNet with attention, weighted focal loss for the rare calcite class, domain-robustness augmentation for scanner shift, and a TFLite export tested for edge inference.
Segmentation
Swin-UNet
Focal loss
TFLite
MGTKAN: KANs for Tabular Data
A Kolmogorov-Arnold network architecture for tabular data, with multi-residual deep graph network feature fusion and knowledge distillation. Benchmarked across clinical and general datasets including glioma grading and Parkinson telemonitoring.
KAN
Graph networks
Knowledge distillation
Contextual AI for Predictive Maintenance
A proof of concept on Google cluster traces testing whether job context predicts machine failures better than KPI telemetry. Context models reached an AUPRC of about 0.70 against about 0.46 for KPI-only baselines, and job priority came out as the single most predictive feature.
LightGBM
VAE
PyTorch
GCP
LLM Competency Extraction at Scale
Labour market intelligence over 4.9 million job postings: K-Means into 1,000 job clusters, a local Llama 3.2 extracting skills, knowledge and abilities, fuzzy mapping onto both O*NET and ESCO, and a knowledge graph of skill transferability between clusters.
LLM
Llama 3.2
Knowledge graph
NetworkX
RL Test Prioritisation
Weighted pairwise covering arrays combined with a Q-learning agent that learns which tests to run first from execution feedback, balancing bugs found, coverage and cost. Ships with a React and Express demo that plots the learning curve against static ordering.
Reinforcement learning
Q-learning
React
Banglar Kotha: Bangla Chatbot
A conversational assistant in Bangla built with RASA and a Bangla spaCy pipeline, served over a REST webhook. Built during the chatbot internship at Brainekt.
RASA
spaCy
Bangla NLP
Digit Recognizer
Machine Learning-based project to detect handwritten Bengali digits using a pre-trained model.
Python
TensorFlow
CNN
Flower Classification
Machine Learning and computer vision-based project to detect flowers using PyTorch and TPU.
PyTorch
Computer Vision
Transfer Learning
Pneumonia Detection
Analyzed chest radiogram images to detect pneumonia using image processing and machine learning.
Medical Imaging
OpenCV
Deep Learning
Vehicle Detection
Deep learning and computer vision approach using Mask-RCNN and EfficientNet.
Mask-RCNN
EfficientNet
Object Detection
Educational Website Development
Developed a learning platform for DevIncept using HTML, CSS, JavaScript, and Django.
Django
JavaScript
Full Stack
Diabetic Retinopathy Blindness Classification
Processed retina dataset and classified using ensemble deep learning model.
Ensemble Learning
Medical AI
Image Processing
Lung and Colon Cancer Detection
Classification of lung and colon cancer using ensemble model (Xception + MobileNet).
Xception
MobileNet
Cancer Detection
Toxic Comment Classification
Deep Learning and NLP-based project to classify harmful comments on social media.
NLP
BERT
Text Classification
