Skip to content

FalconX

AI voice agent that finds the cheapest hotel in a city, phones several hotels in parallel to verify the rate and book, and emails a live-updating confirmation. 2nd prize at the TechTO hackathon.

Private repository. Ask me for the source
Back to all projects
The FalconX repository on GitHub
The private repository on GitHub

Overview

FalconX is an AI voice agent built at the TechTO hackathon on October 17, 2026, where it won 2nd prize. You type a prompt like "cheapest hotel in Ontario, CA for Oct 24–25, email team@…" and FalconX parses it, discovers candidate hotels from several sources, phones them in parallel through an ElevenLabs voice agent to verify the real rate, and emails a confirmation whose body keeps updating as better phone quotes come in. A Next.js "race control" UI streams the whole run live over SSE.

The problem

Every team ran the same prompt at the same moment, and the agent had to actually phone hotels — no browser automation, no ready-made booking agents. Judging was speed first, then price, then distance to downtown. That made it a production problem as much as an AI one: real phone lines, IVR menus, hold queues, flaky pricing APIs, and a hard clock.

What I built

  • A FastAPI backend on Cloud Run exposing /run, /search, /stop, SSE /events, receipt and audio pages, ElevenLabs server tools, and an HMAC-verified post-call webhook.
  • Prompt parsing into a typed intent (city, dates, guests, guest name, email, team) using rules plus Gemini on Vertex AI.
  • Multi-source hotel discovery started at t=0 — a pre-built city database, Google Places, SerpApi Google Hotels, StayAPI, and LiteAPI — de-duplicated by phone number and ranked by total stay price, then distance to downtown.
  • An outbound-call worker pool that dials from a live queue as soon as the first candidates land, using ElevenLabs over a Telnyx SIP trunk, with chain-slot limits, hold caps, and slot release on dead dials.
  • An ElevenLabs agent defined as code (prompt, first message, tools) that navigates IVR menus with DTMF, asks the rate, negotiates, and reports back through typed server tools: report_quote, request_go_ahead, and report_confirmation.
  • An orchestrator that decides BOOK / HOLD / DECLINE on every quote, does the nights-and-tax math server-side, asks for a read-back on implausible prices, and auto-stops 3 minutes after the first verified price (8-minute cap).
  • Live email at the first phone-verified quote: rendered PNG, transcript image, receipt link, and a "listen to the call" link, with a fixed subject line.
  • Durable runs: recordings and run state archived to GCS so email links and ?run= replay links survive restarts, plus a DRY_RUN mode and fake-hotel receptionist agents for rehearsals.
  • A pytest suite of 100+ tests across search, pricing, currency, geo, storage, agent config, and the call backend.

Key technical decisions

Typed tools for decisions, transcripts for display

Live transcripts drive the UI and email, but booking decisions only come from server-tool calls with typed fields — so the agent can get an immediate BOOK or HOLD reply instead of the system guessing prices from partial text.

Built for the clock

Pre-built hotel data, prewarmed pricing caches, dialing from a live queue, front desk before reservations, and reporting the first price immediately before negotiating — every step was tuned for time-to-first-verified-quote.

Production-grade reliability

Cloud Run with an always-on instance, GCS-backed durable state, webhook signature checks, auto-stop rules, retries on IVR keys, and a full dry-run path meant the system behaved predictably under live judging.

Outcome

Won 2nd prize at the TechTO hackathon (October 17, 2026). FalconX placed real outbound calls to hotels, verified rates by phone, and delivered a confirmation email with a call recording and transcript as proof, while the race-control UI showed every call, quote, and decision live.

What I learned

Real phone calls are the least reliable part of any system: IVR menus, hold music, and people who answer mid-sentence. Treating them like any other flaky dependency — timeouts, worker pools, durable state, and clear stop conditions — mattered more than the model prompt. Being ready to change plans (Telnyx instead of Twilio, Gemini instead of Claude, live discovery instead of one static DB) was what got it shipped on time.