Become a great builder
in Data & AI
In-depth writing on machine learning, data engineering, and the systems behind modern AI — for people who want to go from curious to capable.
Most learning stops at the surface.
This one doesn't.
There's no shortage of tutorials that show you what a Transformer does. There's a real shortage of writing that shows you why every design decision was made, and what happens when you change it.
nirmata.dev is written for builders who want that second layer — people who are not satisfied until they could explain it to someone else, implement it from scratch, and debug it when it breaks.
Intuition first
Every concept is explained from first principles before the math. If you don't feel it, you don't own it.
Connections, not silos
Math, code, and systems thinking are taught together — because real problems never respect disciplinary boundaries.
Build to understand
Every major concept comes with runnable code. The fastest path to understanding is making something that works.
Six tracks, one goal: build things that work
Each track goes deep — theory, intuition, and code — so you can go from reading a paper to running experiments.
Deep Learning
Neural networks from scratch, backprop, CNNs, RNNs — and the Transformer architecture that changed everything.
Transformers & LLMs
Attention, positional encodings, pre-training, fine-tuning, RLHF — the complete picture of modern language models.
Data Engineering
Pipelines, warehouses, streaming systems, and the infrastructure that feeds every production ML model.
MLOps
Feature stores, model registries, drift detection, and CI/CD for ML — shipping models that stay reliable.
GPU & Systems
CUDA kernels, memory bandwidth, batching strategies — how to make training and inference fast.
Mathematics of AI
Linear algebra, calculus, probability, and information theory — the tools every serious builder needs.
Start here
Coming soon
Learned Positional Embeddings — BERT's approach
RoPE — Rotary Positional Embedding
Attention from scratch: keys, queries, values
Get new posts in your inbox
No noise. Just a note when there's something new worth reading — a deep dive, a proof, a script to run.