Reading paths

Forty posts is a record. These are the routes.

Reverse-chronological order tells you what happened most recently, not where to begin. Each path below crosses topics and years on purpose, and says why each step comes where it does.

How retrieval actually works

Grounding a language model in documents it was never trained on, from the mechanics up to when not to bother.

You have heard RAG described and want the machinery underneath it.

4 posts · about 11 minutes

1
Are LLMs Actually Thinking?Research
Start with what the model is doing at all: tokens, vectors, attention.
2
How Retrieval-Augmented Generation WorksEvaluation
The loop itself — embedding, cosine similarity, context injection.
3
Document Question-Answering with Pinecone and GeminiEngineering
The same loop as working code, against a real vector database.
4
CAG Over RAG, When Speed Is the ConstraintInference
Then the case against it, when latency is the constraint that matters.

Getting a model into production

Open weights, the hardware they run on, adapting them to your problem, and packaging the result.

You can call an API and now need to run the thing yourself.

4 posts · about 10 minutes

1
Running Google's Gemma Through Hugging FaceEngineering
Get an open model running at all, with the least ceremony.
2
Fine-Tuning Large Language Models: PEFT, LoRA, and QLoRA in PracticeExperiments
Adapt it on one GPU instead of a cluster — PEFT, LoRA, QLoRA.
3
Deploying AI Models with NVIDIA NIM: Production LLMs as MicroservicesInference
Package it as a service rather than a script.
4
LPUs and What They Do That GPUs Do NotInference
And the hardware question underneath all of it.

Agents and the protocol layer

How a model gets access to tools, why that breaks at scale, and where agent-to-agent sits next to it.

You are wiring a model to real tools and want the standards, not the glue code.

4 posts · about 15 minutes

1
Why I'm Excited About MCPSystems
The problem MCP exists to solve.
2
Docker's Dynamic MCP and the Context Window ProblemEngineering
What happens to the context window once you have fifty tools.
3
MCP and A2A Are Layers, Not RivalsSystems
How MCP and A2A stack, rather than compete.
4
Building Teams of AI Agents with CrewAIEngineering
And what you build on top: several agents with roles.

The maths under the models

The two algorithms everything else is built on, and how to tell whether a model learned the right thing.

You want the ground floor before the language models.

3 posts · about 7 minutes

1
Linear Regression ExplainedResearch
Fitting a straight line, worked through properly.
2
Logistic Regression ExplainedResearch
The same machinery turned into a classifier.
3
Explainable AI: Opening the Black Box with SHAP and LIMEResearch
Then how to check a model is right for the right reasons.

Set up a machine you can work on

Choosing the operating system, living with the choice, and fixing the things that break first.

You are early enough that the tooling is still in the way.

5 posts · about 11 minutes

1
Choosing an Operating System for ProgrammingSystems
The choice itself, without a single right answer.
2
Why Programmers Keep Choosing LinuxSystems
Why the tooling keeps pointing one way.
3
Windows to Linux: What Changed in How I WorkSystems
What actually changes day to day after switching.
4
Fixing pip's externally-managed-environment ErrorEngineering
The first error a fresh Python install throws at you.
5
Getting Started with DjangoEngineering
And something real, from empty machine to running app.