Bite-Sized AI

This section is my personal exploration of the AI ecosystem. To feel more comfortable with this topic, I needed to understand the foundations, without going too deep into the mathematics behind them.

I hope these articles can help you understand these concepts as well.

Understand the foundations

How LLMs Work
An overview of how large language models work during inference
Tokenizer
How text is converted into tokens and IDs before reaching the model
Embeddings
How token IDs are converted into vectors that are refined inside the model
Transformer
How transformer layers progressively refine token representations during inference
Attention
How tokens retrieve information from other tokens in the context
MLP
How the MLP transforms each token’s hidden state after attention
KV Cache
How the KV cache helps an LLM avoid recomputing information during decode

Explore the ecosystelm

Hugging Face
Where open-source models live: how to find a model, read a model card, and download what you need
Reading a Model Name
What qwen2.5:7b-instruct-q4_K_M actually means: family, parameter count, variant, and quantization explained
Ollama on macOS
Run open-source LLMs locally using Ollama and add a web UI with Open WebUI
Ollama on Kubernetes
Deploy Ollama with persistent model storage and a web UI, using an initContainer to pull models automatically
vLLM introduction
PagedAttention, continuous batching, and why a GPU is essential for production inference
vLLM on Kubernetes
Install the NVIDIA GPU Operator and serve your first model via the OpenAI-compatible API
RAG (Retrieval-Augmented Generation)
How to make LLMs answer questions about your own data: the indexing pipeline, embeddings, vector stores, and retrieval
Tool Calling
How LLMs can call functions and take actions: defining tools, handling responses, and feeding results back
MCP (Model Context Protocol)
Introduction to the protocol that allows an LLM to access external services

Articles published on Exoscale blog