Juan Andrés Girón Caballero — Writing https://giron.ai/en/notes I write when something took me a while to understand. en Shipping facial recognition on $200 hardware https://giron.ai/en/notes/reconocimiento-facial-hardware-barato https://giron.ai/en/notes/reconocimiento-facial-hardware-barato The model was never the problem. The per-device budget was AI in production Tue, 14 Jul 2026 00:00:00 GMT What an AI agent actually is (and what is hype) https://giron.ai/en/notes/que-es-un-agente-de-ia https://giron.ai/en/notes/que-es-un-agente-de-ia If it cannot fail observably, it is not an agent: it is a demo AI in production Tue, 30 Jun 2026 00:00:00 GMT I built an LLM from scratch: what nobody tells you about transformers https://giron.ai/en/notes/llm-desde-cero https://giron.ai/en/notes/llm-desde-cero I don't use tools I can't derive. So I derived this one Building in public Fri, 12 Jun 2026 00:00:00 GMT RAG in production: what breaks that tutorials never show https://giron.ai/en/notes/rag-en-produccion https://giron.ai/en/notes/rag-en-produccion The most expensive failure mode is the one that doesn't look like a failure AI in production Thu, 28 May 2026 00:00:00 GMT Multi-tenant from day one: the decision that saved me https://giron.ai/en/notes/multi-tenant-dia-uno https://giron.ai/en/notes/multi-tenant-dia-uno It is the one thing you cannot retrofit without rewriting the product SaaS and architecture Sat, 09 May 2026 00:00:00 GMT The mathematics of anti-spoofing, without the hype https://giron.ai/en/notes/matematica-anti-spoofing https://giron.ai/en/notes/matematica-anti-spoofing Telling a face from a photo of that face is a threshold problem, not magic Mathematics without hype Wed, 22 Apr 2026 00:00:00 GMT