The Car Delivery Control Room

Build the place where people and agents work together to get every car delivered, and one real agent for one part of the job.

Materials

Challenge repo: data, brief & starter Participant brief Operations context Participant brief (PDF)Bahn slides (PDF)

Who we are

Bahn delivers used cars for car dealers. When a dealer buys a car, often in another country, one of the freelance drivers in our network collects it from the seller and drives it to the dealer. This year that is over 25,000 cars and more than 1,500 drivers, across Finland, Denmark and Germany, run by a team of about 50.

We deliver in about 2 days. A truck takes 2 to 4 weeks, because it waits until it has enough cars on one route to fill the truck. Our drivers drive each car on its own, so nothing waits. When cars arrive sooner, dealers sell them sooner and make more money.

What our operations team does

Our operations team runs every delivery: phone calls, emails and paperwork, and something changes every hour. The people the team depends on are not our staff: the seller who has the car, the driver, the ferry line, the person who signs at the other end. Any of them can be late, wrong or silent. A normal week has things like these in it:

A driver flies to Germany, misses the train, and cannot make the pickup slot.
The seller said the car was ready. The driver arrives and it is not.
The seller wants a paper nobody mentioned before the car can leave.
The ferry is sold out, so the driver needs a later ship and maybe a hotel.
An order says Tampere in the delivery address and Helsinki in the notes.

None of these is hard on its own. The hard part is that each one starts a chain, and today a person in operations holds the whole chain in their head.

Where we are going

Our agents already do a lot of the routine work: they sort gate codes and papers before a driver turns up, price jobs and offer them to drivers, and spot problems while a car is moving. The next step is agents that do most of the work in the background and bring a person only the cases that need one.

A person should see only the orders that need them, and for each one just enough to decide: what happened, what the agent already found out, what it suggests, and a way to accept it or change it. People make the judgement calls, and the agents do the legwork.

1. The control room. A place where people and agents work together to get every car delivered. It shows the orders that need a person, says why, and keeps the rest out of the way. A person can open any order and see in a few seconds what happened, who said what and when, what the agent has done, and what the next decision is.

2. One real agent for one area. Go deep on one part of the business: an agent that uses tools and that a person can talk to. It does the cheap work first (reading, asking, looking things up), then finishes the job or brings a person a clear decision. Some areas, or propose another one that fits car delivery:

Pickup readiness, by phone. Call the seller or pickup place, confirm the car and papers are ready, deal with phone menus, try again later if nobody answers.
Driver travel. Plan how a driver gets from Helsinki to a private address in small-town Germany and how the car gets home: trains, buses, taxi, ferry, flights, hotels. Stop before any real purchase.
On the road. Breakdowns, missed ships and delays: gather the facts, ask the driver, propose the next step to a person.
Driver support. Handle drivers' routine questions and hand the hard ones to a person with the context already gathered.

The rules of the game

The agent never contacts a customer (anyone who pays Bahn) on its own. Drivers and pickup places are fine to contact.
Start with approvals: a person says yes before the agent changes a booking, a date, a cost or messages a customer. Asking questions and looking things up need no yes.
The agent has to be real: a language model doing the work in a loop. Everything around it can be mocked, including a simulated driver who answers in the demo.
The control room is for people in operations, not engineers. It must not look like a developer tool.
Everything in the packet is made up. Keep it that way: no real people, messages or bookings.

What you get

A fully synthetic but realistic dataset of 100 orders with dealers, customers, drivers, cars, bookings, travel, conversations, alerts, documents and an event history (some of it messy on purpose), 14 example proposals from Bahn's own agents, 58 made-up attachments, an optional Next.js + Postgres starter app, and OPERATIONS-CONTEXT.md on how the work goes. It's all in the GitHub repo below: start with the README.

What to hand in

By Sunday 15:00, add both to your write-up on the hub (the write-up is your submission):

A demo video, two to three minutes. Show and pitch your solution, and show how you built it with coding agents like Claude Code or Codex. A screen recording with your voice is fine.
A public GitHub repository that runs from a fresh clone with the steps in your README. Forking Bahn's repo is the easy way to start. Tests and agent traces are welcome.

Practical details

Any model, API, framework or UI library, local models too. Questions: Bahn's channel on the event Discord, or their booth. Bahn is hiring: come and talk to them.

How it's judged

What Bahn looks for

The control room, 40%. Does it bring a person only the orders that need them, with just enough to decide? Can you see where every fact came from? Does the agent do the cheap work first, know when to stop, and hand over with what it found? When a person decides, is it clear what happens next?

Overall control room8 pts
How did they use coding agents5 pts
TASTE5 pts

How you worked with AI, 30%. How you used coding agents while you built it: what you asked for, what you checked, what you threw away, and where you did not take the first answer. Judgement over the number of tools.

Taste, 30%. Would Bahn's operations people use this every day, without a tutorial? Does one glance tell you what is wrong and what to do, even with hundreds of orders? Does it look made with care: readable text, calm spacing, sensible empty and loading states?

Each judge scores each criterion from 1 to 5.

How to submit

Your write-up is your entry

Strict deadline: Sunday 15:15. No late entries.

01Pick this challengeOn My team — you can pick more than one and change picks until the deadline.
02Write it up on the hubWhat it does, how it works, repo / demo / video links and screenshots. Bahn judges from this write-up.
03Hit “Submit to Bahn”Here or on My team, before the deadline. You can submit the same project to each challenge you picked.

Teams building

29 teams

20€1A_Usual_Dude1Aalrigth2BALI2Claudin' it3Da3mons1Hai la focul lui Sumedru!!!1Ho's Nephews2Hot Dino1It_Worked_On_Mine1J & Sons3KJA3Kia.M1Leo and Leon2MKEP4One1Prometheus3Sell me this hackathon2Ship it & Pray3Shoreline1Solo Warrior1TeamName2Torpedo2Vroom Vroom1Y.E.O.3Zero to One5aBARofchocolate5bob the builder1pepsi4