AI ML Digest

Featured

Choose a metric by what a mistake costs you

Choose a metric by what a mistake costs you

Sep 25, 2026

•

6 min read

What Pydantic actually buys you over a dataclass (and when to reach for it)

What Pydantic actually buys you over a dataclass (and when to reach for it)

Sep 20, 2026

•

6 min read

Building RAG: The Old Way vs. The LCEL Way

Building RAG: The Old Way vs. The LCEL Way

Sep 16, 2026

•

3 min read

All posts

Everything we have published, newest first.

Idempotent tools: making agent retries safe

A timeout doesn't mean the call failed, so build tools that can run twice without doing damage.

Oct 25, 2026

•

2 min read

Approval gates: put them where the risk is

Ask a human only about actions you can't undo, and the approvals will still get read.

Oct 18, 2026

•

4 min read

Control plane vs data plane, the half of a managed service nobody demos

Creating, metering, isolating and upgrading 2,000 vector store clusters is a bigger build than the engine, and here is the order to build it.

Oct 14, 2026

•

9 min read

Where your GPU memory goes: training vs inference

Four buckets, one 8B model, one 24GB card: how to estimate what fits before you launch, and which lever shrinks which bucket.

Oct 9, 2026

•

8 min read

Parameter-efficient fine-tuning compared: what LoRA and its relatives trade off

A memory, quality and complexity comparison of LoRA, QLoRA and (IA)3, plus a rule for picking one under a fixed GPU budget.

Oct 8, 2026

•

6 min read

Load more