Gurpreet Deol

Data Science Bootcamp

Data Science from beginner to advanced level

The MLHub Posts

The Math Behind Modern AI, Derived and Then Coded in NumPy

The maths modern AI leans on, derived step by step and checked against clean NumPy. Dense layers, regression, cross-entropy gradients, backprop, PCA and SVD, Adam, numerical stability and …

Vision Transformers From Scratch, Then Onto the Edge (A Production Guide to ViTs, Quantisation and Jetson-Class Devices)

Build a Vision Transformer from first principles in PyTorch, work out its parameters and FLOPs by hand, then shrink it with ONNX, INT8 and TensorRT so it runs …

Polars Under the Hood: Why It Beats pandas, and How to Use It on Real Pipelines

A first-principles tour of Polars: Arrow memory, SIMD, Rust threads and the lazy query optimiser, then a runnable pandas-versus-Polars pipeline, a migration cheat-sheet, Rust tooling (DuckDB, uv, Ruff, …

Fine-Tune an LLM on One GPU with LoRA and QLoRA From First Principles (Math, NumPy and Production Code)

Learn how LoRA and QLoRA let you fine-tune a 7B model on a single consumer GPU, from the memory arithmetic and low-rank math to NF4 quantisation in NumPy, …

How Agentic AI Really Works Under the Hood: The Agent Loop, Tool Calling and the Model Context Protocol (MCP)

Build an agent from scratch in 60 lines of Python, compare the main agent architectures, watch the raw JSON-RPC messages of the Model Context Protocol, write an MCP …