Roy's Blog

ML experiments, implementations & research notes

~/gourab-roy — bash

❯ cat about.md

Welcome to my tech blog! I write articles on AI and ML concepts that are relevant to both research and machine learning competitions.

Here, I share insights from my coding journey and include code references to help you better understand and implement the concepts discussed.

❯

Articles

// in-depth technical breakdowns & experiments

Meta-Stacking

Ensemble learning where base model out-of-fold predictions feed a meta-learner. Extracts signal that no single model could capture alone.

Ensemble Stacking ML

8 min read

Adversarial Validation

A sanity check for your data pipeline. Train a classifier to distinguish train from test — detect distribution drift before it burns you.

Validation Data Drift CV Strategy

6 min read

Time-Series Forecasting with RNNs

A complete pipeline: rolling window feature engineering on climate data, wired into an RNN for next-day temperature prediction.

RNN Time-Series Deep Learning

10 min read

CatBoostClassifier on Polynomial Features

Feature engineering meets gradient boosting on the Palmer Penguins dataset. Polynomial expansion + CatBoost's native categorical handling — 97.1% test accuracy.

CatBoost Polynomial Features Classification

7 min read

K-Means Clustering with Inertia Measurement

Building K-Means from scratch in pure NumPy and finding the optimal K via the Elbow method — no scikit-learn clustering.

K-Means Clustering Unsupervised

8 min read

Solving CartPole with a Deep Q-Network

Teaching an agent to balance a pole using reinforcement learning and PyTorch. Experience replay, epsilon-greedy exploration, mean reward 226.2.

Reinforcement Learning DQN PyTorch

10 min read

Variational Autoencoder on MNIST

Learning a structured latent space for image generation with PyTorch. Reparameterization trick, KL annealing, and smooth digit generation.

VAE Generative Models Deep Learning

9 min read

Decision Tree Classifier from Scratch

Building Gini-based tree splitting in pure NumPy on the Iris dataset. No black-box: every prediction follows a traceable path of if-else rules.

Decision Tree Gini Impurity From Scratch

7 min read

PCA from Scratch

Dimensionality reduction via SVD in pure NumPy. 10D to 2D, validated against sklearn to 1e-6. Scree plot, loadings heatmap, and reconstruction error analysis.

PCA Dimensionality Reduction NumPy

8 min read

DBSCAN from Scratch

Density-based clustering in pure NumPy. No k required. Core points, border points, noise detection — tested on three Gaussian clusters with random scatter.

DBSCAN Clustering From Scratch

7 min read

Adam Optimizer from Scratch

From plain gradient descent to Adaptive Moment Estimation. Momentum, RMSProp, bias correction — then a benchmark pitting Adam vs SGD vs Momentum on XOR.

Adam Optimization From Scratch

12 min read

Backpropagation from Scratch

Training a neural network with no autograd whatsoever. Forward pass, cross-entropy loss, full backward pass with chain rule, and weight updates written out in plain NumPy.

Backpropagation Neural Networks From Scratch

9 min read

// no articles match

Stay Updated

// get notified when new articles drop

Tech Stack

// tools & frameworks powering the research

Python · C++ · SQL
PyTorch · CNNs · Recommender Systems
OpenCV · Medical Imaging · Reconstruction
SBERT · FAISS · RAG · Cross-Encoders
PySpark · Databricks · Git
Time-Series · RNNs · Forecasting