Skip to content
~/ai-ml-handbook by @ka1manov

ls -lh ~/ai-ml-handbook/articles/

Long reads

Six articles that go deeper than the roadmap has room for. Each one covers a mechanism, the trade-offs it forces, and the failure modes that follow — the things that decide whether a system works in production rather than in a notebook.

articles 6|words ~16k|read ~76m total|by @ka1manov

articles/

Transformers, from attention to KV cache

Scaled dot-product attention and why it is scaled. Multi-head projections as a partition rather than an addition. Positional encoding from absolute to RoPE, and the algebra that makes rotation encode relative position. Then the KV cache, and how it stops being an optimisation and becomes the constraint that decides your batch size.

words ~2.7k|read ~13m|level mid → senior

Where the money goes when you serve a language model

Prefill is compute-bound and decode is memory-bandwidth-bound, and almost every serving mistake comes from conflating them. Continuous batching, paged attention, quantisation, speculative decoding — with a table of which lever moves which metric, because several trade directly against each other.

words ~2.6k|read ~12m|level senior

Fine-tuning: mostly, a guide to not doing it

Fine-tuning teaches form; retrieval supplies facts. Confusing those is the most expensive mistake in applied LLM work. The decision framework, the ladder of cheaper things to try first, LoRA and QLoRA mechanics, DPO against RLHF, and the costs nobody budgets for.

words ~2.7k|read ~13m|level mid → senior

Evaluation engineering

The work that makes an LLM product improvable rather than merely changeable. Building a golden set that is not contaminated, decomposing the pipeline so a regression is attributable, programmatic checks before model-based ones, calibrating a judge so its numbers mean something, and a paired bootstrap before believing a small difference.

words ~2.7k|read ~13m|level mid → senior

Vector search internals

How flat, IVF, product quantisation and HNSW actually work, and which parameter moves which point on the recall–latency–memory triangle. Tuning against a measured target rather than a feeling. And why filtered search breaks ANN indexes, with the three things production systems do about it.

words ~2.5k|read ~12m|level senior

Agents, and why they fail

Per-step success raised to the number of steps is the arithmetic that governs everything, and it explains the whole gap between the demo and the product. Tool design, context management, a failure taxonomy with mechanisms rather than symptoms, the security model, and why the right answer is often to write the sequence instead.

words ~2.7k|read ~13m|level mid → senior