15 exhaustive articles on building real products on top of hosted models: prompting, RAG, agents, evals, cost, safety, and shipping. The applied counterpart to LLMs from Scratch, not a re-run of its theory.
0 of 15 read
Start anytime
You're not training a model, you're renting one, by the token.
A prompt is a spec, not a magic spell.
Getting reliable JSON out of a model that predicts text.
Why 'just paste more text in' fails.
Chunking is where RAG quality is actually won or lost.
The pipeline everyone draws as one box.
Plan, call a tool, observe, repeat.
When one agent stops being enough.
'Memory' is a product decision, not a feature flag.
'It feels better' is not an eval.
A decision framework, not a tutorial.
Token economics is a real budget line, treat it like one.
The attack surface of an LLM app isn't the model, it's everything feeding it.
Debugging something non-deterministic.
Every prior article, tied into one working system.