About

I'm Juan Manuel Infante Quiroga, an AI developer focused on Generative AI, NLP, and AIOps based in Bogotá, Colombia. I build production AI systems for the layer most tutorials stop before reaching. At Finanzauto, I architected a RAG platform that processes hundreds of thousands of WhatsApp conversations daily — 1,000+ concurrent users, automated collections, real-time customer support.

Next Token is the documentation layer for that work — the orchestration decisions, the infrastructure tradeoffs, and the failure modes that don't appear in prompt engineering posts. Long-form articles when an idea earns the depth, field notes when it doesn't, rendered notebooks when the code is the point.

The work centers on four problems that production consistently makes hard.

RAG pipelines are the first. Retrieval sounds mechanical until you're assembling context at scale — deciding what enters the window, in what order, while keeping latency acceptable when hundreds of requests arrive simultaneously.

Orchestration is the second. LangGraph workflows, Twilio-integrated callbots, Airflow pipelines that need to survive restarts: orchestration is where correctness gets expensive. The challenge isn't wiring the pieces together — it's keeping them coherent when one fails.

NLP and multimodal agents are the third. Combining speech recognition with LLM reasoning introduces latency constraints that pure-text systems never face. The decisions at that layer — how to buffer, how to recover from a misrecognition mid-conversation — don't appear in any chatbot tutorial.

AIOps is the fourth. Knowing whether a system that handled 300,000 messages last week will handle next week's is the problem that makes every production deployment honest. Evaluation frameworks, observability pipelines, and the CI/CD discipline that turns a working model into a reliable service.