§05 · index

Writing & notes

Posts are published on Medium. Notes live here — a running reference on how LLMs, transformers, inference, and retrieval systems actually behave, written from first principles and corrected against implementation work.

Notes

N1 · Neural networks · Training · Fundamentals — 6 min
Deep learning, stated precisely
The parts that are easy to get almost-right: what a neuron computes, the difference between backprop and gradient descent, and why stacked layers build a feature hierarchy nobody designed.
N2 · Transformers · Attention · Embeddings — 8 min
The transformer forward pass, token by token
From raw text to a contextual representation: tokenization, embedding lookup, Q/K/V, attention, and why stacked layers refine rather than rewrite.
N3 · Inference · KV cache · Serving — 10 min
What actually happens at inference
The generation loop, why decoding parameters are post-hoc, how the KV cache works, what GQA and quantization trade away, and how one model serves many requests at once.
N4 · RAG · Retrieval · AI systems — 9 min
Building RAG from the boundaries in
Retrieval and grounded generation are two steps, not one. Define the interfaces first — chunking, embeddings, vector store, retriever, generator — and treat every model output as untrusted input.
N5 · Backend · Durable jobs · Reliability — 8 min
Backend patterns for AI work
AI calls are slow, expensive, and unreliable. The backend around them needs durable jobs, worker-owned state, retry-safe writes, and a way to recover when a queue drops a message.

Published · Medium

Everything on Medium ↗

Writing & notes

Deep learning, stated precisely

The transformer forward pass, token by token

What actually happens at inference

Building RAG from the boundaries in

Backend patterns for AI work

Building production-ready LLM applications with RAG

MLOps best practices: from Jupyter to production

Deep learning for medical imaging: explainable pneumonia detection