Portfolio

Projects

A collection of Kaggle competition notebooks, deployed web apps, research reproductions, and open source personal projects.

Competitions

Kaggle Projects

View all on Kaggle

Kaggle

Images Tiles Generator from WSI

Preprocessing pipeline to extract tiles from Whole Slide Images (WSI) for the HuBMAP Kaggle challenge.

Computer VisionWSIPreprocessingSegmentation

Kaggle

Glomeruli Segmentation — SeResNext50 UNet

UNET model with SeResNext-50_32x4d encoder trained with Dice loss for semantic segmentation of Glomeruli areas in kidney tissues.

PyTorchUNetSeResNeXtMedical Imaging

Kaggle

Lung Anomaly Detection — Faster R-CNN

Faster R-CNN trained with IceVision for detection of anomalies in chest X-rays from the VinBigData dataset.

Object DetectionFaster R-CNNIceVisionRadiology

Kaggle

Audio Signal Processing - EDA and Feature Engineering

Exploratory data analysis and feature engineering for audio signal processing tasks.

Audio ProcessingEDAFeature Engineering

Deployed

Web Applications

Live Demo

Phishing URL Detector

Web application for real-time phishing URL detection using a Random Forest classifier trained on NLP features extracted from URL structure and text patterns.

FlaskRandom ForestNLPCybersecurity

Open Source

Libraries & Frameworks

Open Source

HistoCoreML

Production-ready ML framework for computational histology. Covers WSI I/O, patch tiling, Macenko H&E normalisation, TorchScript/ONNX segmentation, foundation model embeddings (UNI, ViT, CTransPath), biomarker extraction, and multi-format output (TIFF, RLE, GeoJSON, Zarr).

PyTorchONNXDigital PathologyWSISegmentationApache-2.0

Experiments

Research & Tutorials

Research

Graph Neural Networks for Fraud Detection

Comparison of GCN and GAT architectures on the Elliptic Bitcoin dataset for illicit transaction classification. GAT improved recall from 0.45 to 0.67.

PyGGCNGATFraud DetectionCrypto

Research

Colorectal Histology Classification

ResNet-50 trained from scratch on the Colorectal Histology MNIST dataset achieving 92% accuracy across 8 tissue types.

ResNetPyTorchDigital PathologyClassification

Research

Metro Traffic Forecasting — LSTM

Comparative study of 1-layer, 2-layer, and bidirectional LSTM architectures for multivariate multi-step traffic volume prediction.

LSTMPyTorchTime SeriesForecasting