Someone organises a business trip and needs to know who is coming, where each person is flying from and when they would rather travel. Normally that is a form, and half the team ignores it. Here they get an email asking them to phone a number instead. A voice agent takes the call, asks the three questions, and the answers land against the trip before anyone has opened a spreadsheet.
The Pipeline
Employee records go in as PDFs. Landing AI parses each one to markdown and then extracts it against a small schema, which pulls out a name, an email, a location and a phone number. Those become employees in the database, attached to a trip.
Resend emails each of them the number to call. The VocalBridge agent answers, confirms whether they are attending, asks which city they are leaving from and whether they would rather travel during the day or at night, and writes all of it back against the trip. Sabre supplies the travel data, and anything genuinely private, the sort of thing nobody should be reading aloud to a voice agent, is collected through a one-time secure link instead.
Four Hours
Built in 4 hours at the DeepLearning.AI hackathon in the Computer History Museum, Silicon Valley, California. First commit 11:05. Last commit 15:07.
Six third-party services in an afternoon only works if the two halves never wait on each other. So the front end was built against a fake. One module chooses between a mock implementation and the real one behind a single interface, picked by an environment variable, and no page in the app knows which it is talking to. The screens could be built and clicked through while the routes underneath them did not exist yet, and when the real ones landed the switch was one variable rather than a rewrite.
That is the part worth taking away. Under a deadline the instinct is to write less, and what actually saved time was writing one extra layer that did nothing useful on its own.
What It Isn't
It is not deployed. The host it ran on during the hackathon has been retired, and the backend has no authentication of any kind, which was acceptable for four hours on a private address and is not acceptable in public. What remains is proper auth on the server, tests, and a deployment that does not hand out API keys.
It was a two-person project. My brother Ildar built the server and every integration in it. I built the client, set up the VocalBridge agent that takes the calls, and tested the pipeline end to end.