Writing
Occasional long-form pieces on agentic AI, enterprise systems, and building in LATAM. Substantive over frequent.
We Swapped the Brain of a Production AI Agent and Measured Whether Its Personality Survived
On July 8, 2026 we changed the LLM behind Francesca, San Roque's WhatsApp receptionist, from Claude Opus 4.5 to Grok 4.3 with a byte-identical persona prompt — and measured what survived. Six months of production data on what a persona spec actually controls, and what a model switch silently rewrites.
What 27,000 WhatsApp Conversations Taught Me About Production AI Agents
Nine and a half months of production data from Tere, the AI sales agent we built for Colombian fabric retailer Telas Real. The architecture decisions that worked, the ones that were hard, and why grounded database lookups beat model knowledge.
Human-in-the-loop for voice agents: why your architecture decides whether it's possible
Which voice agent architecture you pick determines whether human review is even possible before delivery. This piece breaks down cascaded vs native speech-to-speech tradeoffs for enterprise compliance.