Alcancía LLC · 2024–2025

Noah

Live

AI customer support agent built for telecom MVNOs. First deployed with a small MVNO in Mexico — zero to production in ~4 months, no prior playbook. I owned the full build cycle.

The problem

Customer support across small telecom MVNOs ran entirely on human agents — billing inquiries, service requests, network complaints, all handled manually with no AI layer. No playbook existed for what this could look like. The brief was open: build something that works in production, not a pilot.

What I built

A multi-agent conversational AI system deployed on messaging. A supervisor agent classifies every inbound message and routes it to a specialized sub-agent — billing, service requests, technical issues, or escalation. Each sub-agent has its own context window and toolset. For technical issues, the agent communicates directly with the customer's phone to pull live diagnostics — line state, connection health, hardware identifiers — instead of asking the customer to read them out.

Key decisions

Hub-and-spoke over linear chain

Inbound queries were too variable to route deterministically through a linear pipeline. A supervisor model that classifies intent and dispatches to specialists gave flexibility to add sub-agents without restructuring the whole system.

Define what the agent should not handle

The most important product decision was the refusal boundary. I scoped out edge cases that generate liability — legal disputes, account terminations, payment disputes — and routed them directly to human agents. This drove a meaningful reduction in escalation rate.

Persistent state across sessions

Using LangGraph's native state management, each conversation maintains context across messages. The agent remembers prior interactions within a session window — returning users don't re-explain their issue.

Direct device interrogation for technical issues

Rather than asking the customer to describe their network state, the agent pulls diagnostics from the device itself — signal strength, connection state, hardware identifiers. The customer talks to the agent in plain language; the agent talks to the phone in protocol. The triage step disappears.

Outcome

AI customer support agent deployed in production with a Mexican MVNO. Zero to production in ~4 months — no prior playbook, no external consultants, no pre-built vendor solution. The agent handles live customer interactions today.

Want to go deeper on Noah?