DEV Community: Vavus AI

Visit Carmo de Minas with a translator

VavusAI — Fri, 24 Jul 2026 15:19:51 +0000

Visit Carmo de Minas with a translator

Carmo de Minas is the kind of place a traveler can pass through too quickly.

It is a small town in Minas Gerais, inside the Mantiqueira de Minas coffee region, and it does not behave like the obvious Brazil itinerary. The point is not a crowded beach or a famous skyline. The point is the slower thing: coffee fields, mountain weather, small cafes, local roads, farm visits, and people who can tell you what the place means if you can stay in the conversation.

That is the travel use case for translation.

You can visit Carmo de Minas and order coffee without speaking much Portuguese. But if you want to understand the locals, you need more than pointing and smiling. You need to ask where the beans came from, understand when someone explains processing, ask whether a cafe sells the beans it serves, follow directions, read signs, and hear what locals actually recommend.

Use Vavus Keyboard before you arrive: WhatsApp, Instagram DMs, booking notes, directions, and thank-you follow-ups in Brazilian Portuguese.

Use Vavus AI when you are there: local conversations, translated calls, camera translation for menus/signs/labels/receipts, and help with what to say next.

The goal is not to perform perfect Portuguese. The goal is to show that you care enough to ask clearly and listen.

Carmo de Minas is small, but the place gets bigger when translation lets you understand the people inside it.

Full guide: https://vavusai.com/blog/visit-carmo-de-minas-with-translator

Business angle: https://vavusai.com/blog/carmo-de-minas-coffee-translation-travel

Learn more at https://vavusai.com.

Phone Calls With Family Who Speak Another Language

VavusAI — Fri, 24 Jul 2026 15:19:06 +0000

Your father calls every Sunday at eleven his time. You already know how this goes, because it's gone the same way for years. "Bueno." "Hola, Papá." "¿Cómo estás?" "Bien, bien, ¿y tú?" "Bien, bien." A pause you've both learned to expect. "¿Y los niños?" "Bien, comiendo mucho." Another pause, then a small laugh at nothing in particular, almost in unison. "Bueno... te quiero, hijo." "Yo también, Papá." The call ends at one minute fifty, same as last Sunday. Nothing was really said. Everything that could be said, in the ten phrases you both fully trust, was.

This is the call between the big calls — not the birthday video chat with the whole family crowded around a laptop, not the holiday check-in with translated captions running underneath everyone's face. Just an ordinary Tuesday or Sunday, a plain voice call, the same phrases in the same order.

Video calls, phone calls, and texts aren't the same problem

A quick way to see where each tool actually fits:

Moment	What's actually happening	What helps
The birthday video call, whole family on screen	Occasional, high-visibility, everyone taking turns	Live translated captions on a video call
The ordinary Sunday phone call	Frequent, low-stakes, voice only, no camera	Real-time call translation inside the Vavus AI app
The "did you eat?" text between calls	Constant, tiny, easy to skip	Vavus Keyboard translating as you type

The phone call is the middle one, and it's the one most families never bother fixing — because a ninety-second call doesn't feel broken, it just feels like that's how it is.

What actually happens on the call

Stated precisely, because precision matters here: inside the Vavus AI app, you can place or receive a call where each person sets their own language, and the app translates what's said in each direction as the call happens — one side speaks Spanish, the other hears English, and back again. That's real-time translation on a call inside the app. It doesn't reach into the ordinary cellular network, and it doesn't intercept or translate a call placed through a regular phone dialer or carrier. If the call isn't happening inside Vavus AI, Vavus isn't part of it. Smaller claim than "translates your phone calls" — the accurate one.

The days in between

Vavus Keyboard covers the texting layer: type in your language, it sends in theirs, their reply comes back in yours, inside whatever messaging app is already installed. A photo of dinner with three words under it. A twelve-second voice memo. It doesn't replace the call. It fills the six days around it.

The honest limit

None of this produces fluency. It won't catch a joke on the first beat — humor runs on timing, and translation adds a half-second of lag a joke often can't survive. An idiom about a dog and its tail may cross over as a sentence about an actual dog. That's not a bug to fix in a future release. It's what translation is: a workaround for the years of immersion it actually takes to think in someone else's language, not a substitute for them.

Weigh that against the alternative, which for most families isn't "fluent conversation" — it's "bien, bien," repeated because it's safe, on a two-minute call, forever. A translated call with rough edges, where an actual sentence of real information crosses the line, is worth more than a fluent-sounding call where nothing did. Participation beats silence.

The bottom line

A two-minute call where you actually exchanged something real beats a two-minute call where you both just performed warmth at each other. Try it on the next call — set each person's language in Vavus AI and see what six minutes with your own father sounds like instead of two.

Originally published at vavusai.com.

How to Translate a Document — Or Just a Photo of One

VavusAI — Fri, 24 Jul 2026 15:18:21 +0000

You're in a landlord's office with a lease in your hands, and clause 14 is the one that matters — something about renewal terms, or who pays for what — and you can't quite work it out. Or a school sends home a dense notice about a schedule change. Or you're at the pharmacy counter with a medicine box, trying to work out how many times a day and with or without food.

None of these are emergencies. That's exactly why they're easy to put off. But paper (and paper-like problems: a photographed form, a screenshot, a label) shows up constantly, and most people don't know there's a fast way through it. There are really two different jobs here, and they call for two different tools — plus one job neither of them should ever claim to do.

Job one: point your camera at it

If it's a sign, a menu, a label, or a single page — something you can photograph in one frame — this is the fast path. In Vavus AI, it's Image OCR Translation:

Open it from the Translate hub, and point your camera at the text or upload a photo/screenshot you already have.
Snap it. The app reads the text out of the picture (OCR — optical character recognition) and detects the language.
Get translated text back, right next to what it found.
Copy, save, or act on it.

Job two: when it's a real document

A lease isn't one frame. Neither is a contract, a multi-page form, or a letter where the layout carries meaning. Photographing five separate pages loses the structure that makes a document a document — which paragraph belongs to which clause, which line is a total versus a subtotal.

That's what Document Translation is for: upload the file — a PDF, a Word document, a scanned contract, a medical intake form — and it comes back fully translated with the original layout intact.

The three-way comparison

	Photo translation	Document translation	Certified human translation
Best for	Signs, menus, labels, single pages	Contracts, forms, multi-page files	Immigration paperwork, court filings, anything "certified" required
Preserves layout?	No — just extracts the text	Yes — structure, tables, formatting	Yes, plus a signed statement of accuracy
Speed	Seconds	Minutes (whole file)	Days, human-reviewed
Legally certifiable?	No	No	Yes
Good for	Understanding a sign or label right now	Understanding a real document, keeping its structure	Anything an institution requires as "certified"

What to double-check afterward, honestly

Anything OCR-derived — text pulled from a photo or scan rather than typed — deserves a second look. A blurry photo or cramped handwriting can trip up character recognition in ways typed text never does:

Sanity-check every number: a dose, a rent amount, a due date.
Double-check names and proper nouns — the words most likely to get slightly wrong.
Reread dates in full, especially where day/month order could flip the meaning.
If something looks off, compare the translation against the original line by line.

The honesty note that matters most

Translation is not the same thing as a certified translation. If a document has legal or official weight — immigration paperwork, a court filing, a birth or marriage certificate, anything an institution specifically requires as a "certified translation" — you need a certified human translator, full stop. A certified translation comes with a signed statement of accuracy, sometimes a notarization; the institution reviewing your paperwork checks for that specific credential, not for whether the translation merely reads correctly. A tool like this is excellent for understanding what a document says before you bring it to a professional — it is not a substitute for that professional once the paperwork itself needs to be certifiable.

Not sure whether your situation needs a certified translation? Ask the institution requesting the document.

One line on privacy

Documents and photos you translate are handled with client-side encryption where applicable, so you're not putting your lease or medical form through a black box.

FAQ

How do I translate a photo of text?
Open Image OCR Translation in Vavus AI, point your camera at the text or upload a photo/screenshot, and it reads and translates on the spot.

Is this a certified translation?
No. It's for understanding what a document says, not for producing a legally certifiable translated document. For immigration paperwork, court documents, or anything an institution requires "certified," use a certified human translator.

What should I double-check afterward?
Numbers, dates, and names — especially anything pulled from a photo or scan, since OCR can occasionally misread a character.

The bottom line: paper still shows up everywhere, and most of it needs the right tool, not an app that only handles one sentence at a time. Snap a photo for a sign or single page, upload the file for a real document, double-check the numbers and names either way — and bring in a certified human translator when the paperwork itself needs to be certified.

Originally published at vavusai.com.

How to Clean Up a Rough Message Before You Send It

VavusAI — Fri, 24 Jul 2026 15:17:31 +0000

You know the draft. Maybe it's the apology to a friend after you bailed on dinner twice in a month, and every version sounds either too casual or too much like a hostage note. Maybe it's the ask to your boss for a schedule change, buried under three layers of "sorry to bother you." Maybe it's the complaint to a landlord about heat that's been out for four days, rewritten because it sounds too angry, then rewritten again because it sounds too apologetic for asking at all.

The pattern's always the same: type it, read it back, wince, delete a line, retype it — and by the fifth pass you're not editing for clarity anymore, you're just anxious. The facts haven't changed across six phrasings. The problem was never the facts. It was the tone.

Why tone breaks down in text

In person, tone rides on things you never consciously register — pitch, pacing, a pause before the hard part, a half-smile that says you're not actually mad. None of that survives being typed into a box. All that's left is the words, and words alone are a thinner signal than most people assume while writing them.

Communication researchers Justin Kruger and Nicholas Epley, studying email, found that senders routinely overestimate how accurately readers pick up their intended tone — including sarcasm — because writers "hear" their own intended tone in their head while typing, a cue the reader never gets. The message feels obviously warm to whoever wrote it and can land as flat or curt to whoever reads it, especially if that person's already stressed or on edge. Rereading your own draft doesn't catch this, because you already know what you meant.

The manual checklist

Shorter sentences. Qualifiers are where tone gets muddy — say the thing, then stop.
One ask, not three. A message with three requests buried in it reads as pushy, because the reader can't tell which one matters.
Cut the hedges. "Sorry to bother you but," "this might be a dumb question" — reads as insincere padding after the hundredth time. State the ask; let tone carry the politeness.
Read it out loud. The single best free tool. Tone that looks fine on screen often sounds sharp in your own voice.

Yourself vs. a friend vs. AI cleanup

	Rewriting it yourself	Asking a friend to read it	Vavus Keyboard AI cleanup
Speed	Slow — you reread your own bias	Depends if they're free	Seconds
Catches tone you can't hear in your own head	Rarely (you're anchored to intent)	Usually yes	Yes
Fixes grammar for a second-language writer	No	Sometimes, if they're patient	Yes, same pass
Available at 11pm before you hit send	Yes, but you're the weak link	No	Yes
Changes what you're actually saying	No	Occasionally (their opinion creeps in)	No — same facts, same ask, different delivery

Where AI cleanup fits, precisely

Vavus Keyboard's rewrite and tone tools take the message you already wrote and adjust how it sounds — tighter, warmer, more direct — without changing what you're saying or adding claims that weren't there. It's not a ghostwriter. It won't invent a more diplomatic version of events or turn a complaint into a compliment. It rewrites for delivery, not intent. You're still the author.

The second-language layer

If you're writing in a language that isn't your first, there's a second problem stacked on the tone problem: slightly-off grammar or phrasing reads as curt or hard to parse to a stranger, not as "still learning." A tool that fixes phrasing and tone in the same pass handles both — clarity and delivery — at once.

Send it once

Run the checklist first: short sentences, one ask, no hedging, read it aloud. When that's not enough, or the language isn't your first, cleanup tools give the draft a pass without putting words in your mouth.

Originally published at vavusai.com.

How to Travel Without a Phrasebook (and Still Talk to Everyone)

VavusAI — Fri, 24 Jul 2026 15:16:40 +0000

You're at a ticket counter in a train station in Kraków, thumbing through the "Getting Around" chapter of a phrasebook looking for something close to "which platform for Warsaw," while the line behind you shifts its weight. You find a phrase, say it in your best approximation of the phonetic spelling, and the clerk answers back fast, three sentences, none of them in the book. You catch nothing.

That's the actual failure mode of a phrasebook: it's a one-way script for a two-way event. It prepares you to say something and does nothing for what comes back — and travel is almost entirely made of what comes back.

The three kinds of things you actually need to say

Transactional — buying, ordering, paying, asking how much. Short, high-frequency, needs a real answer back.
Navigational — finding your way, asking for help when something's gone sideways. Needs speed and clarity.
Social — small talk, thanks, apologies, compliments. Not required, but it's where the warmest exchanges of a trip happen.

A phrasebook tries to cover all three with the same static text. Not enough for how any of them actually play out.

What replaces it

Tool	Covers	Needs a connection?	Language coverage
Paper phrasebook	One-way scripted phrases you say	No	Whatever's printed, no listening
Vavus AI conversation mode	Live two-way spoken translation	Yes — data or WiFi	100+ languages (voice)
Offline language pack	Type-and-translate text, no signal	No	Same text coverage, no live voice
Vavus Keyboard	Typed messages — hosts, drivers, front desk	Yes	200+ languages (text)

Two honest notes worth being precise about. Conversation mode needs a connection — it's not a magic trick that works with zero signal. And Vavus AI's coverage is layered: text translation spans 200+ languages, but live voice — the speak-and-listen kind — covers a narrower band, roughly 100+ languages. Check your destination language's voice coverage before relying on it as your only plan.

For the typing parts

A lot of trip friction isn't spoken — it's typed. A message to an Airbnb host about a broken lock, a text to a driver running late, a note to front desk. Vavus Keyboard translates as you type inside whatever app you're already using (WhatsApp, iMessage, a booking app's chat), and translates what comes back. The messages that go wrong on a trip are usually the short, rushed, typed ones — getting those right the first time saves the afternoon.

What's actually worth prepping before you fly

Download the offline language pack. Real and worth doing — but it's text only, not live conversation. It's your backstop for a dead zone on a train, not a replacement for conversation mode.
Learn five to ten phrases yourself. Hello, please, thank you, excuse me, sorry, how much, where is, do you speak English, cheers, goodbye. Saying "thank you" badly in someone's own language does something an app can't — it shows you tried first.
Skip the paper phrasebook. It prepares you to talk, not to be talked to — and travel is mostly the second one.

Where this doesn't apply

None of this replaces a real interpreter for anything high-stakes — a medical situation abroad, a police matter, a legal document. Use a certified human interpreter or your embassy there. For the everyday texture of a trip — buying, asking, thanking, chatting — this is exactly the territory these tools are built for.

The bottom line

Live voice translation handles the conversation as it happens, a translating keyboard handles the messages in between, an offline pack is a smart backup for dead zones, and a handful of memorized words still buys you something no app can: a moment of genuine effort with a stranger. Bring all four, and leave the paper book at home.

Originally published at vavusai.com.

Translation is travel UX: six Tokyo rules the interface does not explain

VavusAI — Fri, 10 Jul 2026 09:40:28 +0000

Tokyo is full of excellent interfaces: colored rail lines, numbered station exits, floor markers, ticket machines, pictograms, and carefully structured queues.

The hard part for a visitor is not always missing information. It is understanding which information is carrying social meaning.

A floor marker is not decoration; it is the start of the queue. A silent carriage is not an invitation to take a call. The absence of a public bin does not mean a wrapper belongs beside the vending machine. A translated menu may explain the food while a handwritten notice explains the ordering system.

Six small rules make the interface legible:

Keep transit conversations quiet and move phone calls elsewhere.
Follow platform markers and let passengers exit first.
Carry rubbish until an appropriate bin appears.
Keep cash available even when cards and IC payments work most of the time.
Do not apply a default tipping workflow.
Translate context before acting on an assumption.

This is why travel translation should be treated as a UX layer rather than a phrasebook.

Camera translation turns visual state into context. Live translation turns uncertainty into a specific question. Keyboard translation handles the asynchronous flow before and after the trip: booking, confirmation, clarification, and follow-up.

The system still needs human awareness. Translation cannot tell you every unwritten norm. It can, however, reduce the number of moments where a traveler guesses because asking feels too difficult.

The best travel interface is not the one that makes a visitor feel clever. It is the one that helps the visitor create less friction for everyone else.

Full guide and official sources:
https://vavusai.com/blog/tokyo-travel-etiquette-rules

Translation as a travel workflow: Carmo de Minas coffee

VavusAI — Thu, 09 Jul 2026 08:27:23 +0000

This post is not about a generic translator box. It is about a workflow.

The use case: a traveler or small business operator visits Carmo de Minas, a coffee town in Brazil's Mantiqueira de Minas region. They want to do more than order a drink. They want to talk to a cafe, message a producer, call to confirm a farm visit, ask about lots and processing, understand labels, and follow up afterward in Brazilian Portuguese.

That creates several surfaces:

Typed messages in WhatsApp, Instagram DMs, email, booking apps, and notes.
Live voice conversations in cafes, cooperative offices, cars, and farm visits.
Calls when someone answers with audio instead of text.
Camera translation for menus, labels, receipts, coffee bags, invoices, signs, and tasting notes.
Follow-up messages where tone matters because the business relationship is still forming.

The product split is useful:

Vavus Keyboard handles the typed layer: draft, translate, reverse-translate, rewrite, and send from the app where the conversation already lives.

Vavus AI handles live conversations, calls, camera translation, documents, and richer context.

The lesson is that travel translation is not one action. It is a chain: message before arrival, speak while there, translate the physical world, and follow up afterward.

Full guide: https://vavusai.com/blog/carmo-de-minas-coffee-translation-travel

The best travel translation setup for expats and travelers

VavusAI — Wed, 08 Jul 2026 11:16:29 +0000

Most travel translation workflows fail because they assume every problem is a text box.

It is not.

Some travel problems are typed:

host messages
driver replies
booking app chats
emails
notes
forms
social posts

Those belong in the keyboard. Vavus Keyboard brings dictation, translation, reverse translation, rewrite, and AI agents into the text field itself.

Other travel problems are spoken, visual, or contextual:

talking with locals
menus
station signs
tickets
PDFs
calls
"what should I say here?"

Those belong in the main app. Vavus AI handles voice, camera, documents, calls, and richer context.

The practical architecture is simple:

Surface	Best for
Vavus Keyboard	typed messages, dictation, translation, rewrite, reverse translation, quick agentic help
Vavus AI	locals, menus, signs, documents, calls, camera, fuller context

This matters because copy-paste is the hidden tax in translation products. Every time a user has to leave the app where the work is happening, paste into a translator, copy back, then fix tone, the workflow gets weaker.

The better pattern is: keep the keyboard close to the text field, and keep the main app responsible for bigger multimodal moments.

Full guide: https://vavusai.com/blog/best-travel-translation-setup-expats-travelers

Working on a team that chats in a language you're still learning

VavusAI — Tue, 07 Jul 2026 17:40:48 +0000

It is 9:14 a.m. and the #eng-standup channel is already twelve messages deep. Someone posted a decision about the deploy window, someone else pushed back in three quick lines, and by the time you have re-read the second message to catch a word you did not recognize, the conversation has moved on. You could ask - the team is genuinely patient - but you already asked Priya to slow down twice this week, and you do not want today to be the third time.

This is what a lot of remote and global teams look like now. The working language on Slack or Teams often is not anyone's first language in the room - it is just the language everyone agreed to use, moving at the speed of whoever is most comfortable in it. As more teams hire across borders and time zones, that gap between "the language we chat in" and "the language I am fluent in" is quietly common, even on teams nobody would call broken.

The usual workarounds are thin: ask someone to repeat themselves, switch to a language you both know better, or nod along and piece it together afterward. All of it works, sort of, for a while - until you are the one always asking, always half a step behind, always last to weigh in because you were still translating the last three messages in your head.

Reading: keeping pace with the channel on your own

The channel does not wait. New messages land while you are still working through the last one, and falling behind means scrolling back through a thread you already half-read, or asking someone to summarize what you missed.

Vavus Keyboard translates as you read. A message, a thread, a long project update - read it inline in your language right there in Slack, Teams, email, or whatever the team actually uses. No second app, no switching windows, no losing your place. You keep pace the way everyone else does: by reading it, as it happens.

Writing: knowing what your message actually says before you hit send

Reading is half the problem. The other half shows up when you go to reply and second-guess yourself - did that come out right, does it sound as clumsy in their language as it feels in your head, is there a word in there that means something you did not intend.

This is where reverse translation earns its keep. Type your reply in your own language, or your best attempt at the team's, and Vavus Keyboard shows what it actually says when translated - before you send it, not after someone reacts to it strangely. You catch the wrong word or the sentence that reads more abruptly than you meant, while you can still fix it, instead of sending a message and quietly worrying about it for ten minutes.

Meetings: captions that keep you in the room

Chat is one thing. A live standup, an all-hands, or a client call at full native-speaker speed is another - there is no re-reading a sentence someone already said three sentences ago.

For spoken meetings, Vavus AI provides live translated captions: you read what is being said, translated into your language, as close to real time as the conversation allows. You follow the discussion instead of catching every fourth sentence and guessing at the rest, and you can jump in with a question without needing the whole meeting to pause for you.

One honest limit: live captions are for following along and participating in the room. They are not a substitute for a professional interpreter on anything contract-level, legally binding, or otherwise high-stakes in that meeting - a negotiation, a formal review, a decision with legal weight. For the normal run of standups, syncs, and all-hands, captions do exactly the job you need. For the meeting where the wording genuinely matters legally, bring in a certified interpreter for that specific conversation.

The cost of always being the one who asks to slow down

Here is the part that does not show up in any retro: asking a colleague to slow down, repeat something, or switch languages is free the first time, fine the second. By the fifth or sixth, even a genuinely patient, welcoming team starts to feel like something you are imposing on - not because anyone said so, but because you are the one keeping count. Some people respond by going quiet: fewer questions in meetings, waiting for the notes instead; fewer messages in the channel until they are sure the wording is right. None of it is about ability - it is about not wanting to need extra patience one more time, even from a team that would happily give it.

Self-serve translation removes that math entirely. You are not asking anyone for anything - you are reading a translated message, checking your own reply, or following a caption, on your own time. The team does not do less for you; you just stop needing them to. That is the shift that actually changes how someone participates - more questions asked, more ideas posted, less waiting to be sure before speaking up.

Where this earns its keep

Distributed engineering and product teams: standups, PR reviews, and design discussions in a shared working language nobody in the room grew up speaking.
Sales and customer success teams: internal deal notes and Slack threads in the company's language, on top of customer conversations in a different one entirely.
New hires ramping up: the first few months on a team chat are the hardest to keep pace with - this closes the gap early instead of waiting for it to close on its own.
Contractors on international teams: joining a client's Slack for a few weeks is no reason to fall behind on every thread.

What it costs

Vavus Keyboard is $14.97/month on web ($14.99 on Apple) for unlimited dictation and translation, or pay-as-you-go with tokens - reading, writing, and reverse-checking every message where you already work. For meetings, Vavus AI starts at $9.97/month on web ($9.99 on Apple) for live translated captions. Against months of quietly falling behind on your own team, both are an easy trade.

Frequently asked questions

How do I keep up with a team chat that moves fast in a language I am still learning?

Use a translation keyboard on the channel itself. Vavus Keyboard translates incoming messages as you read them, inside Slack, Teams, or whatever your team uses, so you keep pace without asking a colleague to slow down or summarize.

Can I check my own message before I send it?

Yes. Type your reply and Vavus Keyboard shows a reverse translation - what it actually says in the team's language - before you send it, so you can catch a wrong word or an awkward phrase while you can still fix it.

Do live captions work for meetings, not just chat?

Yes. Vavus AI provides live translated captions for spoken meetings - standups, all-hands, client calls - so you can follow along and participate in real time.

Are live captions a substitute for a professional interpreter?

Not for anything contract-level, legally binding, or high-stakes - bring in a certified interpreter for that conversation. For the normal run of team meetings, live captions are exactly the tool you need.

The bottom line: a team chat moving in a language you are still learning is not a sign you do not belong. It is friction - and a translation keyboard for reading and writing, plus live captions for meetings, quietly remove it. You read at the pace everyone else reads, send messages you are actually confident in, and follow meetings without leaning on the same patient colleague every week. Try it at vavusai.com.

Originally published at vavusai.com.

Ordering Food Abroad When You Can't Read the Menu

VavusAI — Mon, 06 Jul 2026 16:32:56 +0000

Everyone who has traveled has done it: standing in front of a menu board covered in a script you can't read, pointing at whatever looks least alarming. Sometimes it works out. Sometimes what arrives is a plate of something you genuinely did not want.

"Point and hope" is a strategy built entirely on luck. Here's what actually replaces it — and the one part where luck is not an acceptable substitute.

Three ways to order, compared

Approach	What it gets you	Where it fails
Pointing at the menu	Fast, no effort, sometimes fine	Pure guessing — no idea what's coming, no way to ask a question
Photo-translating the menu	You read and choose a real dish, know what you're ordering	Doesn't answer live questions like spice level, substitutions, recommendations
Asking the server via voice translation	Real back-and-forth: recommendations, spice level, ingredients, modifications	Needs a data or WiFi connection to work in real time

Read the menu instead of guessing at it

Vavus AI's camera translation lets you point your phone at a menu board, a printed sheet, or a chalkboard special and see it translated in front of you — a menu you can actually read and choose from, not word-by-word gibberish. Most of what goes wrong with point-and-pray ordering isn't complicated; it's simply not knowing what a dish is.

Ask what a photo can't answer

A translated menu tells you what a dish is called. It doesn't tell you how spicy the server's "mild" really is, or whether the kitchen can leave out an ingredient. Vavus AI's conversation mode handles that: you speak, the server hears their language, they answer, you hear yours — a real conversation instead of a static page.

Where this stops being a guessing game: allergies

This is the part that matters most, so it's not a footnote. Organizations like FARE (Food Allergy Research & Education) track how many people in the U.S. alone live with a food allergy — a meaningful share of the population, and for some, a reaction is life-threatening, not just uncomfortable.

A translation app is a helpful backup, not the primary safety tool. For anyone with a severe allergy, a professionally translated allergy card — stating the specific allergy clearly in the local language, ideally reviewed by a professional, handed directly to kitchen staff — should be doing the heavy lifting. It's unambiguous, doesn't need signal or battery, and removes the risk of a mistranslated word when the stakes are severe. Voice and photo translation are genuinely useful around the edges of an allergy-aware trip; they support that plan, they don't replace it.

A few small etiquette notes

Asking your server what they'd recommend is often preferred over a long list of specific requests.
Portion sizes and course structures vary more than a translated menu suggests.
A little effort with "please" and "thank you" tends to get warmer service than a flawless translated order with no eye contact.

FAQ

Is a translation app safe to use for food allergies?
It's a useful backup, not a substitute for the real safety tool. A professionally translated allergy card, ideally reviewed by a professional, handed directly to kitchen staff, is the standard practice for anyone with a severe allergy.

What if the restaurant has no signal?
Live voice translation needs a connection. Photo translation can work with limited connectivity depending on setup. A printed allergy card works with zero signal, which is one more reason it belongs in your pocket regardless.

The bottom line: read the menu instead of guessing at it, ask your server the questions a photo can't answer, and if a real allergy is involved, carry a professionally translated allergy card as your primary safeguard.

Originally published at https://vavusai.com/blog/ordering-food-in-another-language

Agentic keyboard AI: from text field to next action

VavusAI — Sat, 04 Jul 2026 16:06:33 +0000

Originally published at vavusai.com.

Most keyboards try to predict the next word. Vavus Keyboard is moving in a different direction: it helps with the next action around the words already on screen.

The new agentic keyboard makes the existing Ask AI and Voice AI surfaces more capable. You can ask for a rewrite, a reply, a translation, a summary, a live answer, or a quick check without leaving the app where you are typing. The keyboard stays a keyboard. It does not become a full workspace, a calendar client, or a document manager. It gives you a fast action layer inside the text field.

That distinction matters. When you are answering a message, filling a form, writing a support reply, or translating a sentence in a mobile app, the expensive part is often not the model. It is the context switch. Copy the text. Open another app. Paste. Ask. Copy the answer. Return. Paste again. Fix the tone. Repeat.

Vavus agentic keyboard is built to reduce that loop.

Selected text first, clipboard when needed

The new mobile keyboard workflow starts with the clearest user signal: selected text. If you highlight text in the field and tap a keyboard action, Vavus uses that selection first.

If there is no selected text, the keyboard can still use the clipboard as the fallback. That keeps the older copy-paste workflow working for cases where the app does not expose selected text cleanly, or where the user intentionally copied something from another app.

This is a small product decision with a big trust benefit. A text document may contain far more than the sentence you want to transform. The keyboard should not treat the whole document as the prompt. It should act on the selected passage when you choose one, and only fall back to copied text when there is no selection.

For AI context, the keyboard can include a small window around selected text: a few words before and after the selection. That helps the model understand whether "it" means the customer order, the meeting, or the document, without sending the whole field.

What agentic means here

Agentic does not mean the keyboard gets permission to do everything. In the keyboard, it means Ask AI and Voice AI can work through a short, read-only tool loop before writing the final answer back into your text field.

Examples:

Reply with context: Select a customer message and say, "Reply politely and mention I can help tomorrow."
Use live information: Ask, "What is the weather in Bucharest tomorrow, and turn it into a short reply?"
Check a public page: Paste or select a URL and ask the keyboard to summarize the page in plain language.
Translate with intent: Select a paragraph, translate it, and keep the tone professional.
Rewrite in place: Ask for a softer, shorter, more formal, or more direct version of the selected text.
Quick lookup: Ask for time, currency, units, weather, stock quotes, or a short web-backed answer while staying in the current app.

The important part is not that the keyboard has a chat panel. It is that the result lands where the work is already happening.

Ask AI and Voice AI use the same idea

Text AI is useful when you want to type an instruction: "make this warmer" or "translate this into Spanish for a client."

Voice AI is useful when typing the instruction is slower than saying it: "Answer this and say I am free Thursday afternoon." The keyboard can turn your spoken instruction into an agentic run, use the selected text as context, and return a draft you can review before sending.

That makes voice less like dictation alone and more like a command layer for writing. You are not only speaking the text. You are telling the keyboard what to do with the text.

Light by design

There is a reason the agentic keyboard is intentionally light.

A keyboard lives inside other apps. It should be fast, focused, and careful. The keyboard does not need every tool from the full Vavus AI app. It does not need to connect calendars, change documents, send messages on your behalf, or manage projects from a tiny typing surface.

The keyboard tool set is read-only and practical: search, fetch a public page, translate, detect language, list languages, get time, weather, stock quotes, currency, and unit conversion. That is enough for the quick language and writing jobs people actually do from a keyboard.

The heavier agentic workflows belong in the main Vavus AI app: files, documents, account connections, richer assistants, approvals, and longer-running work. The keyboard is for the moment when you are already inside WhatsApp, Gmail, Slack, Notes, a browser form, a CRM field, or a patient intake message and need the next sentence to be right.

Standalone keyboard and Vavus AI keyboard

This update matters for both keyboard surfaces:

Vavus Keyboard: the standalone keyboard app for mobile and desktop users who want dictation, translation, reverse translation, Ask AI, Voice AI, and language tools in the apps where they type.
The Vavus AI built-in keyboard: the keyboard inside the main Vavus AI mobile app, so users who start in the broader app still get the same mobile text-field workflow.

The goal is consistency. A user should not have to learn a different behavior depending on which Vavus keyboard surface they installed. Select text first. Use the clipboard only when there is no selection. Let AI use a small context window when it helps. Review the result before sending.

Why this changes the feel of writing

The old translation loop is mechanical. It asks you to manage text movement. The agentic keyboard asks what outcome you want.

Instead of:

copy the message,
open a translator,
paste,
translate,
copy the result,
paste it back,
rewrite the tone,
check the meaning,

you can stay in the field and ask for the result directly.

That is useful for everyday writing, but it is especially useful when language and tone both matter:

A support agent needs to be clear and calm.
A salesperson needs a follow-up to sound natural.
A traveler needs a quick reply in the local language.
A bilingual family member wants a message that keeps the original warmth.
A clinic front desk needs approved administrative text to be understandable and reviewed.
A student needs help turning a rough note into a clean message.

The keyboard does not remove judgment. It removes friction before the judgment step.

Privacy and review still matter

Agentic keyboard AI is powerful because it sits close to user text. That is also why the rules are strict.

The selected passage is the primary input. Clipboard fallback exists for explicit copy workflows. Field context is limited. The keyboard actor uses read-only tools. Public web search and page fetch are not the place for private patient, legal, financial, or account-sensitive text.

For high-stakes work, review the result. For healthcare, legal, billing, safety, contracts, or clinical communication, follow the approved workflow and escalation rules. Vavus can reduce the language burden, but it should not remove human responsibility where the stakes are high.

The product direction

The agentic keyboard is not a bigger keyboard for its own sake. It is a quieter one.

It keeps the interface familiar: Ask AI, Voice AI, translation, reverse translation, dictation, and the text field. The difference is what those actions can do behind the scenes before the answer appears.

The promise is simple: when the next step is a sentence, a reply, a translation, a lookup, or a small rewrite, the keyboard should help you do it without leaving where you are.

FAQ

Is the agentic keyboard a chatbot?

Not exactly. It can answer like a chatbot when that is useful, but the job is different. The output is meant to become text in the active field, not a long conversation in a separate app.

Does it read everything I type?

No. The mobile workflow prioritizes selected text, falls back to clipboard text when there is no selection, and uses only a small context window around selected text for AI context. It does not need the whole document to help with a selected sentence.

Can the keyboard use live information?

Yes, for light read-only tasks such as web search, public page fetch, weather, time, stocks, currency, and unit conversion, depending on the user's access mode and available tools.

Is this the same as the full Vavus AI assistant?

No. The keyboard is intentionally light and focused on the text field. The full Vavus AI app is where larger workflows belong: documents, files, account-connected tools, richer assistants, and longer tasks.

Does this replace translation and reverse translation?

No. Translation and reverse translation remain core keyboard actions. Agentic AI adds a more flexible layer for cases where the user wants instruction-following, context, live lookup, or a rewritten reply.

Building translation into the whole communication workflow

VavusAI — Fri, 03 Jul 2026 16:48:37 +0000

Most translation products are designed around one narrow moment: paste text, get text back.

That helps, but it is not how multilingual work actually happens.

A real day looks more like this:

a support call where two people need to talk in real time
a voice note that needs to become searchable text
a screenshot or document that needs OCR before translation
a chat thread where each message may need context from the one before it
a keyboard field inside another app where switching tools breaks the flow

That is the product problem we are working on with Vavus AI: translation as a workflow layer, not a single text box.

Why one-mode translation breaks down

For developers, the hard part is not only translating strings. It is preserving context across surfaces.

A sentence in a support chat, a phrase inside a contract screenshot, and a live spoken reply during a call all have different latency, formatting, privacy, and UX requirements. If the product treats them as the same job, the user feels the cracks immediately.

So our approach is to bring the common translation surfaces into one account:

live voice translation
translated voice and video calls
encrypted messaging with per-message translation
document and image translation with OCR
dictation and transcription
a translating keyboard for the apps people already use
web, desktop, iOS, and Android access

The goal is simple: keep people in the conversation instead of sending them through copy-paste loops.

The architecture lesson

When translation moves across voice, documents, messaging, and keyboard input, a few product constraints become non-negotiable:

Latency has to match the mode

A translated call needs a different speed/quality tradeoff than a document. Voice and calls need quick turns. Longer text can afford more context. The UI has to make those modes feel intentional instead of forcing one pipeline everywhere.

Context matters more than isolated sentences

Literal sentence-by-sentence output is often not enough. Tone, terminology, and intent matter, especially in work, healthcare, travel, and customer support contexts.

Privacy cannot be an afterthought

Translation often contains private material: medical conversations, contracts, family calls, customer issues, internal notes. Vavus AI is built privacy-first with encrypted history, end-to-end encrypted messaging, and HIPAA-ready healthcare accounts for teams that need stricter controls.

The keyboard is part of the product surface

A lot of communication happens outside your app. If translation only works after the user opens a separate screen, it misses the most common workflow: typing where the conversation already is.

What we shipped

Vavus AI is live as an all-in-one translation app for 200+ languages across mobile, web, and desktop.

It includes voice, calls, messages, documents/OCR, transcription, dictation, and a translating keyboard. The short version is: every way to translate, in one app.

Try it here: vavusai.com

I would be especially interested in feedback from developers building multilingual products, voice UX, healthcare workflows, or productivity tools. The core question we keep coming back to is: where should translation live so the user does not have to think about it?