Feature
Every resident reads the answer in their own language, with times and links intact
Sunday, 23:10. An exchange student writes in English that the heating in her room is not working; ten minutes later, the student down the corridor sends a voice note in Italian asking what time the front door closes. Both get the answer in their own language, with the time exactly as the centre has written it. The handbook, the house rules and the FAQs stay in Spanish, which is how the team wrote them: Bedel reads them in Spanish, answers in the language of whoever asks, and its alerts to the team go out in Spanish.
What a resident writes, and which language Bedel answers in
| What the resident writes | What Bedel does |
|---|---|
| "Hi, the heating in my room isn't working" at 23:10 | It opens the ticket just as if it had been written in Spanish and answers in English. The alert to Maintenance does not go through translation: it goes out in Spanish, like every alert to the team |
| "¿A qué hora cierra la puerta?" and then "And on Saturdays?" | It switches language with her. The last message decides, not the first one in the conversation |
| "Ok", "502" or an emoji | A message that does not reveal its language stays in the last one detected in that conversation; if there is none, in the centre's |
| A voice note in French | It transcribes it if the centre uses the default model and the file is 10 MB or less, leaves "🎤 Audio transcrito" in the thread next to the original audio, and answers in French |
| A message in Portuguese that gets an answer half in Spanish | Before it goes out, the safety net spots it and translates it with times, links, figures and names masked. If anything does not match on the way back, the original text goes out |
| Their first message to the centre's number, in German | The notice that they are talking to an AI assistant reaches them in German, with *HUMANO* and *PRIVACIDAD* written like that, because those are the words the system recognises |
| "Do you have rooms available from September?", from a number outside the centre | They receive the same admissions message the centre wrote, translated, with links and email addresses intact. The information PDF goes out as it was uploaded |
The last message decides
Bedel detects the language of the message that has just arrived and answers in it. If the conversation started in Spanish and carries on in English, it switches in that same turn. The context in front of it (handbook, house rules, menu, admission requirement) is in Spanish because the centre wrote it: it is the data, not the language of the answer. Translating it is not changing it, and the data is never changed.
- On every turn, the code gives the model a deterministic hint about the language of the message; the model can overrule it.
- A message with no language of its own ("ok", a room number) stays in the last language of the conversation.
- The handbook and the FAQs are written once, in Spanish; there is no version per language to maintain.
- Six languages are tested: Spanish, English, French, Italian, German and Portuguese.
What is never translated
A clumsy sentence is annoying; a changed time or a broken link does real harm. So there is a closed list of things that go out exactly as they are, whatever the language of the answer, and that list is the same for the model that answers and for the net that translates afterwards.
- Times, dates and figures: if the handbook says "Cena 20:30-22:00", 20:30-22:00 goes out in any language.
- Links, email addresses and phone numbers.
- Proper names: the centre's, the rooms' and people's.
- The words people have to type to the system, *HUMANO*, *PRIVACIDAD* and *menú*, which stay in Spanish, with their meaning explained alongside if needed.
The safety net: if an answer slips out in Spanish
Every answer to someone writing to the centre, whether from the model or from a fixed flow, goes through the same gate before it leaves. If the person does not write in Spanish and the text is not entirely in their language, it is translated. Before translating, the data is replaced with markers and, on the way back, each marker is checked to appear exactly once. If any step fails, the original goes out.
- It only acts when there is a leak in front of it: if the answer is already in their language, it does nothing.
- It reads the text in stretches of eight words, to catch the sentence that starts in English and ends in Spanish too.
- It masks times, dates, figures, links, email addresses, keywords and the names the system knows: the person's, their room and the centre's.
- If it takes more than five seconds, if a marker is missing or duplicated, or if the result still reads as Spanish, the original text goes out and the attempt is left in the system log.
Voice notes
A voice note is turned into text in the language it was spoken in and, from there, follows the same path as a written message: it can open a ticket, cancel a meal or ask a question. The first answer after an audio mentions that writing it avoids mistakes. The original audio is kept in the thread next to what was heard.
- It is transcribed with the centre's default model. If the centre chooses another one, the audio is not quietly sent to a different provider: it goes to the team for review.
- A limit of 10 MB per note.
- If it cannot be transcribed, the person receives "🎤 Ahora mismo no puedo interpretar audios…" and the team gets an alert to listen to it in case it is urgent.
- For someone who has asked to restrict the processing of their data, nothing is transcribed: a person listens to their audio.
The team works in Spanish
Translation is for the people who write to the centre, not for the centre. The alerts, copies and summaries Bedel sends to the team do not go through the safety net and go out in Spanish. What the team does see is which languages people are writing in.
- The centre's metrics show the languages detected in the period, and the nightly summary email counts them too. More in dashboard and metrics.
- The dashboard can be used in Spanish or in English, depending on each team member's preference.
- The AI notice is written in nine languages: Spanish, English, French, German, Italian, Portuguese, Catalan, Galician and Basque. In any other, it goes out in Spanish with English underneath.
- To write to everyone at once there are announcements, which go out exactly as the team drafts them.
What it does not do, said before the demo
- It does not translate what the team writes through the direct chat with a resident: it arrives exactly as the member of staff typed it.
- It does not translate announcements or broadcasts: they go out in the language they are drafted in.
- It does not translate the documents the centre attaches: the admissions PDF or an event poster go out as they were uploaded.
- It does not transcribe audio if the centre does not use the default model, nor voice notes over 10 MB.
- It does not translate the keywords: *HUMANO* and *PRIVACIDAD* are written like that in any language, and the notice explains it.
Frequently asked questions
Do we have to translate the centre's handbook into other languages?
No. The handbook, the house rules and the FAQs are written once, in Spanish. Bedel reads them as data and answers in the language of whoever asks, without changing timetables, rules or figures. If a rule changes, it is changed in one place and the answer in English or French changes with it.
What if the translation changes a time or breaks a link?
The translator never sees them. Before translating, times, dates, figures, links, email addresses, keywords and the names the system knows are replaced with markers. On the way back, each marker is checked to be there exactly once. If one is missing, one is duplicated, it takes too long or the result is still in Spanish, the original text is sent.
Which languages does it answer in?
In the language of the last message. The tested languages are Spanish, English, French, Italian, German and Portuguese. The notice saying it is an AI assistant is written in nine languages, those six plus Catalan, Galician and Basque; in any other, it goes out in Spanish with English underneath.
Does it understand voice notes?
Yes, if the centre uses the default model. The note is transcribed in the language it was spoken in, up to 10 MB, and stays in the thread next to the audio. If it cannot be transcribed, the person is asked to write it down and the team receives an alert to listen to it in case it is urgent.
Does the team need to speak other languages?
Not for what Bedel answers: the alerts to the team go out in Spanish. There are two exceptions worth knowing beforehand. What a member of staff writes through the direct chat with a resident arrives as it is, untranslated, and announcements go out in the language they are drafted in.
How does someone who does not speak Spanish ask to talk to a person?
The AI notice tells them, in their language, to type HUMANO. The word is left in Spanish on purpose, because it is the one the system recognises; if it were translated, they would type a different word that the system does not recognise. When a resident types it and the centre has human support switched on, the alert reaches the phone of the post that handles requests to speak to a person.
Related features
- AnnouncementsThe same notice to many people in the centre, from the dashboard or with '/difundir' on WhatsApp, knowing beforehand how many it reaches.
- Escalation to peopleAnything about money, health, exceptions or conflicts goes to one specific phone by org chart and shift, and emergencies go to the crisis posts.
For each type of centre
- Colegios mayoresFor colegios mayores: the dining hall with its cancellations, packed lunches and allergens; activities with their call and their survey; families and applicants; and what belongs to management, to management.
- University residencesFor residences owned by or affiliated with a university: support in each resident's language, repairs and cleaning with outside contractors, the arrivals of each semester and enquiries about places.
- Student residencesFor student residences and PBSA operators: WhatsApp support in the resident's language, repairs with an owner and a deadline, arrivals, presence and parcels.
- ColivingFor coliving and flex living: support in each resident's language, repairs with an owner and a deadline, common-room bookings, parcels and announcements without groups.
- Researcher residencesFor researcher, graduate and visiting-faculty residences: arrivals any week of the year, replies in each person's language, contracted maintenance and cleaning, and the end of every stay.
- Hostels and alberguesFor hostels, albergues and pilgrim hostels: house questions answered in each guest's language, repairs with an owner and a deadline, and the night covered with nobody at the desk.
- Holiday apartmentsFor holiday apartments, aparthotels and mid-stay lets: replies in the guest's language, repairs with a technician and a deadline, housekeeping requests between stays and house rules.
- Build to rentFor build-to-rent and flex living buildings: repairs with an owner and a deadline even when maintenance is outsourced, parcels with a notice, shared-space bookings, building notices and house rules answered.