Artificial intelligence
Stop eyeballing your prompts
Changing a prompt without measuring it is guessing. Precision, recall, F1 and threshold: the four ideas that turn "looks better" into a number you can compare.
4 min read
All articles
What building with LLMs teaches you.
Notes taken during the work: evaluating a prompt with something other than your eyes, seeing what a model actually does, choosing what to send it, structuring an agent. What worked, what turned out expensive, and why.
7 articles
Artificial intelligence
Context engineering: how to optimise what you send to an LLM
The context window cannot hold everything. Priority per message, token budget, extractive compression and cleanup: the pipeline that decides what stays.
8 min read
Artificial intelligence
I built an AI SaaS without knowing what my LLM was really doing
Luny-AI worked, but I had no visibility between the prompt and the result. What traces, a cost per generation and a golden set would have shown me.
7 min read
Artificial intelligence
ReAct explained for devs: the pattern behind AI agents
Reason, act, observe, loop. The pattern that powers LangChain, LangGraph and CrewAI, and why the LLM never executes the actions itself.
7 min read
Artificial intelligence
LangGraph for JS devs: graphs and nodes without the headache
Graph, Node, Edge, State: four scary words that fit inside a cooking recipe. Explained through a rap battle between Molière and Robespierre.
5 min read
Artificial intelligence
Jaccard similarity: the simplest similarity metric in AI
One division between the intersection and the union of two sets. TypeScript implementation, fuzzy matching with n-grams, and the cases where it falls short.
4 min read
Artificial intelligence
Reaching 400+ AI models with a single API
OpenRouter and Together AI: one key to try GPT, Claude, Llama or Mistral, automatic fallback when a provider goes down, and cheaper open-source models.
3 min read