GlamCode Autopilot Agent

Inspiration

I run GlamCode OS, a real SaaS platform for beauty salon management in Colombia. Talking to salon owners every day, the same pattern kept showing up: WhatsApp booking is slow, inconsistent, and completely reactive — nobody follows up with clients who've gone quiet. When the Qwen Cloud hackathon opened Track 4 (Autopilot Agent), it was the perfect excuse to build the piece I always wanted for GlamCode OS: an agent that doesn't just answer messages, but actually runs the conversation and the follow-up end to end.

What it does

GlamCode Autopilot Agent handles a beauty salon's customer interaction autonomously:

Understands and books — classifies what the client wants (booking, rescheduling, cancelling, a quote, a beauty question, or a complaint) and resolves it against real availability and pricing, remembering context across the conversation.
Never guesses on what matters — complaints and anything ambiguous get escalated to a human review queue instead of an improvised reply. A reviewer approves or discards each case from a dashboard.
Works proactively — a scheduled job finds clients who haven't visited in a configurable number of days and drafts a personalized re-engagement message. Nothing sends without human approval.

How I built it

The backend is Python/Flask, with two agent flows (agente_reservas.py for reactive conversation, agente_proactivo.py for scheduled outreach) sharing a human-review module (revision_humana.py) backed by SQLite. Qwen Cloud (qwen-plus) handles intent classification, strict service/slot matching, and message generation through an OpenAI-compatible API. The frontend is Next.js 16 with a chat simulator, a proactive-outreach view, and a review dashboard. The backend is deployed on an Alibaba Cloud ECS instance running as a persistent systemd service.

Challenges I ran into

Getting the agent to escalate correctly took more iteration than I expected — an early version would confidently quote a service that wasn't even in the catalog just because it shared a keyword with a real one. I rebuilt the matching to be strict and Qwen-driven rather than substring-based, and added real conversation memory so follow-up messages aren't treated as isolated turns. Setting up Alibaba Cloud for the first time (security groups, key pairs, systemd) was also a steep but useful learning curve.

What I learned

That a genuinely useful autopilot agent isn't about maximizing automation — it's about knowing precisely where automation should stop and hand off to a person, and building that handoff to be trustworthy rather than a fallback error message.