Our first live deployment answers calls for a Ukrainian internet provider. Here is what went wrong, what we fixed and what we would do the same way again.

Demos are easy. A real phone line is not. People call from noisy kitchens, switch languages mid-sentence and hang up when they are put on hold. We knew that in theory. The first weeks of the rollout made it very concrete.

Start with the boring calls

Most calls to an internet provider are about the same few things: the connection is down, a payment did not go through, a router needs a restart. We let the agent handle those first and sent everything else to a person. That kept the risk low and gave us real data quickly.

Mistakes we made

We underestimated how often callers interrupt. Early versions waited too long before answering and sometimes talked over people (why full-duplex matters). Fixing turn-taking did more for callers than any change to the language model.

We also handed calls to staff without enough context, so callers had to repeat themselves. Now the agent passes a short summary with every transfer.

The best test was simple: did the caller have to say the same thing twice?

What stays the same

Everything runs inside a private perimeter, recordings follow a written retention rule, and we keep a log of every change. We will keep publishing those notes here.

European Portuguese is not Brazilian Portuguese, and most voice models treat it as an afterthought. That made it the right first language for us.

We are a Portuguese company, so the choice was partly personal. But there is a practical reason too. Our work on multilingual voice systems showed that many European languages get much less full-duplex development than the big global ones.

European Portuguese is a good test. There is real data available, native speakers around us and a clear gap in quality. If our method works here, we can check whether it works for the next language.

Where the data comes from

We train on varied, real pt-PT speech: podcasts, public TV and radio, and parliament sessions. Each source sounds different, and that variety matters more than the raw number of hours.

What comes next

More training, structured evaluation and a demo people can try. After that, a second European language. The steps are on our roadmap.

The prototype works. Now the slowest part of our research is waiting: about a week per training run on four GPUs.

Our current setup uses four GPUs. A big training cycle takes about a week. That was enough to prove the pipeline works, but it is too slow for efficient research.

Every change to the data, the training strategy or the evaluation can cost days. Most of that time is spent waiting, not thinking.

What more compute would change

With 8 or more GPUs we could compare checkpoints side by side, repeat promising runs to separate real gains from noise, and spend more iterations on how the model sounds in Portuguese.

We do not expect linear scaling. We expect fewer days lost between decisions.

This is why we are applying to European compute and R&D programmes (see the roadmap). If you run an HPC centre or a speech research group, we would like to talk.

Running rent.moto.pt means real bookings, real payments and real people waiting at the door. It changed how we build for clients.

When you run your own product (ours is rent.moto.pt), nobody else gets the 7am call. A booking that fails, a stuck payment, a bike that is not where the system says it is: these become your problems, fast.

Plan for the day after launch

Launch day is the easy part. We now plan logs, alerts and admin tools from the first sprint, because we have needed them ourselves.

Small details add up

Clear confirmation emails, a simple way to change dates, one phone number that actually answers. None of this is hard to build. All of it matters to the person renting the bike.

The same lessons go into auto.moto.pt and into every project we build for clients.

Most voice assistants take turns like a walkie-talkie. Full-duplex means listening and talking at the same time, like a normal phone call.

With a typical voice assistant you speak, it waits, then it answers. If you interrupt, it either ignores you or stops completely. That feels robotic.

A full-duplex model hears you while it is speaking. It can stop when you cut in, react to a short “yes” or “uh-huh”, and pick the conversation back up naturally.

Why it matters on the phone

On a support line people interrupt all the time. They correct a number, add a detail or say “no, that is not it”. Handling that well is the difference between a helpful call and a frustrating one.

Our European Portuguese prototype works in full-duplex today, and you can read how our first phone-line deployment went. Training and evaluation are ongoing.