When a traveler books a hotel through a third-party travel platform, they assume the moment they pay, they're confirmed. In practice, that's often not how it works. The reservation gets submitted to the hotel, but confirmation — an actual confirmation number from the property — can take time, sometimes a lot of it. The hotel hasn't confirmed yet. The traveler doesn't know that. They think they're booked.
That gap creates a significant customer service problem at scale. Travel platforms run call centers staffed with agents whose job is to manually call hotels, navigate their phone systems, reach a human, and secure confirmation numbers for pending reservations — one call at a time. It's expensive, labor-intensive, and entirely repetitive. The work doesn't require judgment. It just requires someone to make the call.
This was a 48-hour hackathon engagement. I joined as AI Strategist on a small cross-functional team. The goal: validate whether the manual confirmation call process could be automated with AI, and build a proof of concept strong enough to make the case for a production investment.
I started by interviewing stakeholders from the client's call center operations — the people closest to the problem. I mapped the workflow in detail, identified the failure modes, and defined what success looked like. My co-strategist and I then built the business case together: what KPIs mattered, what the current process cost, and what a working automated system would need to do.
We also designed the evaluation framework — including LangFuse for observability and LLM-as-judge techniques for assessing system output quality. I authored the initial orchestrator prompt that drove the multi-agent system.
Then the team built it. The POC was a multi-agent orchestration system designed to fully automate the outbound confirmation call workflow:
Orchestrator — a coordinating agent managing overall flow and routing between downstream agents. Intent agent — determines the purpose of the call; in this case, a single intent: retrieve a confirmation number. Call agents — a set of specialized agents handling each stage of the actual call: initiating the connection, delivering the greeting, navigating IVR phone menus to reach a live person, conducting the conversation with full guest context, and either securing the confirmation number or exiting with a documented unsuccessful resolution.
The model stack matched each task to the right capability: Claude Haiku for orchestration, Claude Sonnet for the conversational layer, and AWS Nova Sonic for the voice interface itself.
I facilitated the final demo presentation back to the client's stakeholders at the close of the engagement. In 48 hours, the POC was functional.
The proof of concept was delivered at the end of the hackathon. Following the engagement, the client took it into a production build phase. My role was the research, the business case, the evaluation framework, and the orchestrator design — I wasn't part of the production build that followed.
The work that tends to look small — a 48-hour sprint, a single proof of concept — is often the work that decides whether something gets built at all. This one did.
No pitch. No deck. Just a conversation about your problem and whether I can help.