Machine learning papers, translated into code you can break on purpose.

def attention(query, keys, values):
    scores  = [dot(query, k) for k in keys]
    weights = softmax(scores)
    return sum(w * v for w, v in zip(weights, values))

That's the beating heart of Attention Is All You Need. Three lines. Every paper here gets the same treatment: the idea as a function you could set a breakpoint in, real output from running it, and a short list of things to break so you can watch what fails.

ai-ml/

• Attention Is All You Need Vaswani et al., 2017

  Attention is a dictionary lookup with fuzzy keys — everything else is plumbing. Runnable NumPy, four demos that prove it by breaking it.

  English தமிழ் हिंदी

• KV caching and the generation loop Not written yet

  Why inference is affordable. A direct consequence of the causal mask that nobody explains properly.

• LoRA: low-rank adaptation Not written yet

  Fine-tuning without touching the base weights.

Who writes this

Srinivasan Nagaraja Rao. AI engineer. I ship code and models for a living, and I'm learning this field in public — including the parts where the maths scares me.

I'm not a researcher. I'm a programmer. When I read a machine learning paper, I don't want the proof — I want to know what it would look like as a function I could put a breakpoint in. So that's how I write these up.

Why "for dummies"

Because most paper explainers are written by people who are comfortable with the maths, aimed at people heading toward research. That's a real audience. It just isn't mine.

Mine is the engineer who has shipped models, done the MLOps, and still feels like a fraud when someone opens a paper. I'm writing these from inside that exact confusion, while I still remember what was confusing.

The method: translate the idea into code you already understand, then deliberately break it and watch what fails. You debug for a living. Reverse-engineering from broken behaviour is faster than re-reading section 3.2 for the fifth time.