Research & Engineering
A curated portfolio of production code, open-source libraries, and peer-reviewed algorithmic implementations.
AutoClimDS represents a high-complexity breakthrough in autonomous climate intelligence, bridging the gap between fragmented petabyte-scale repositories and actionable scientific discovery. By architecting a machine-interpretable Climate Knowledge Graph that unifies NASA, NOAA, and CMIP6 repositories, the system empowers a multi-agent ensemble to navigate the entire research lifecycle — autonomously reproducing complex, peer-reviewed scientific figures from natural-language instructions alone.
Comprehensive benchmarking toolkit evaluating INT8/INT4 PTQ (Post-Training Quantization) and QAT on small language models (Phi-3, Llama 3.2, Gemma) across ONNX Runtime and Qualcomm NPUs.
Automated, minimalist curve-fitting engine designed to instantly test, score, and select optimal non-linear and polynomial regression models for continuous datasets. Published in Wiley.
Detected vehicles using a Car/Not-Car classifier with heatmap. Calculated inter-vehicle distance using YAD2K (90% Keras / 10% TensorFlow YOLO_v2 implementation).
Recommended ingredients and recipes from a food image. One network identifies ingredients; another devises a recipe. Trained on the Food-101 dataset inspired by Facebook AI research.
Identified duplicate questions on Quora. Advanced NLP feature extraction, EDA, fuzzy features, Logistic Regression, Linear SVM with hyperparameter tuning, and WordCloud generation.
Predicted user film ratings minimising RMSE. Used sparse matrix, user similarity matrix, dimensionality reduction and Matrix Factorisation techniques.
Predicted heart disease from the Cleveland Database (14 attributes). EDA, Decision Tree, Random Forest, Naïve Bayes, KNN comparison with Precision, Recall, F-score.
Semantic segmentation using a convolutional Encoder-Decoder architecture (VGG-16 encoder). Decoder mirrors encoder layers with data augmentation.
Predicted human activities (Walking, Sitting, Standing, Laying) from smartphone accelerometer and gyroscope sensor data using Logistic Regression, Decision Tree and Random Forest.
Solved mathematical equations from images. Collected operators, combined with MNIST in HDF5 format. Used Caffe, TensorFlow and an Abstract Syntax Tree to evaluate results.
Predicted probability of datapoints belonging to 9 malware classes. EDA, feature extraction, multivariate analysis, KNN, Logistic Regression, Random Forest, XGBoost with RandomSearch tuning.
Analysed factors leading to employee attrition. EDA, feature selection, SMOTE, ANN predictions with One-Hot Encoding, evaluating Precision and related metrics.
Classified songs as Hip-Hop or Rock using track features like energy, acousticness and tempo. Applied PCA, Logistic Regression, Decision Trees and resampling techniques.
Detected 32 brand logos (Adidas, Apple etc.) using Flickrlogos-32 dataset. Built a LeNet-5 network with edge detection, morphological processing and YOLOV2.
Early Alzheimer's prediction using SVM, Logistic Regression and Naïve Bayes. Hyperparameters tuned via 5-fold CV with PCA feature selection and model ensembling.
Extracted unique users per month, calculated total bookings, amount spent and room nights stayed. Merged summarised datasets for comprehensive exploratory data analysis.
Banking web application enabling seamless digital transactions: new account creation, unique account numbers, profile editing and inter-account money transfer. Database Management Systems project.
Have an algorithmic challenge, research idea, or want to discuss Edge AI deployment? Let's connect.