ML / AI · Visual reads · Est. 2023

Least
Squares

Visual breakdowns of modern machine learning systems — architectures, system design, and ML design, drawn from first principles for engineers who build things.

AI ML Systems Design Architecture Inference Training Agents RAG Vectors AI ML Systems Design Architecture Inference Training Agents RAG Vectors AI ML Systems Design Architecture Inference Training Agents RAG Vectors

What we cover

Three pillars.
One mission — make hard systems legible.

Architectures

Transformer internals, diffusion stacks, MoE routing — diagrammed from first principles.

System Design

Inference pipelines, vector DBs, distributed training — how production ML actually scales.

ML Design

End-to-end patterns for ranking, recommendations, RAG, and agentic workflows.

Visual reads

View all →

Speculative Decoding in LLMs

Architecture

Speculative Decoding in LLMs

How a cheap draft model and a fast verification pass can deliver full-quality output at a fraction of the latency.

Inference at scale system diagram

System Design

Inference at scale

Batching strategies, KV caching, continuous batching, and the hard economics of serving large models to real users.

The hidden costs of RAG in production

ML Design

The hidden costs of RAG in production

Latency trade-offs between dense retrieval, semantic caching, and long-context models.

YouTube

Watch the systems
come apart.