Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 7 min read

Orivra: The Keyboard That Types My Way

I love Gboard. I want to say that first, because it's true. So when I say my own keyboard, Orivra, replaced it as my daily keyboard, that says a lot already. Gboard has a huge team and years of my typing data behind it.

Orivra: The Keyboard That Types My Way

I love Gboard. I want to say that first, because it's true.

So when I say my own keyboard, Orivra, replaced it as my daily keyboard, that says a lot already. Gboard has a huge team and years of my typing data behind it. Orivra has me, working 13-hour shifts outdoors in Houston, building it on my Galaxy S26 Ultra in between everything else. This was never a desk project with free weekends β€” it got built in the gaps of a working life, early mornings and late nights, one build at a time.

This keyboard's for me. I don't really have plans to let anybody really use it β€” I want a keyboard for me that types my way and is just smart my way.

That's the whole reason it exists. And this is how it got built.

Orivra is personal and closed β€” it's not available for download, there's no public repo, and I'm sharing the story here, not the code or the packs.

How I actually type

I type fast in bursts, sometimes whole phrases with no spaces. I hit the wrong keys. I delete and retype instead of tapping the little suggestion chips. And I've always been a swipe typer β€” a sloppy one.

Most keyboards were built for careful typists and then patched to deal with people like me. Orivra was built sloppy-first. That's my normal, not an edge case.

My bar was simple from the start: when I slow down and tap, it should save me keystrokes. Predictions should chain well enough that I can tap out full sentences. And the strip should never sit there with just 0 or 1 lonely suggestion β€” I want at least 2 real next-word options.

That made the start of Orivra rough, in a funny way. Orivra didn't have swipe at first, so I had to learn to use a normal tap keyboard again, before we even added swipe. It was like going back to the stone ages. πŸ˜‚ All that forced tapping is a big part of why the prediction got as good as it did β€” but when swipe finally landed, I flipped right back to swiping within hours. That's just how I've always typed.

I built a team to build it

I compile every official build myself in AndroidIDE on my S26 Ultra. I'm the owner, and I'm the one who decides what counts as done.

Around me I have a team with actual jobs:

  • Muse (Mu) β€” predictions, language packs, typing smoothness, specs, reviews, and keeping the board so I don't have to
  • ChatGPT β€” lead integration developer, build commands and compiling guidance
  • Claude β€” independent auditor. He checks work he didn't build

We have a rule for what we're allowed to claim. Something is either proposed, source-reviewed, host-tested, Android-compiled, or device-verified. If it wasn't tested on my phone, we don't say it works on my phone.

That rule matters, because fixing one thing and breaking another is the real job. Building the first version was the easy part.

Prediction is the keyboard

Keys are easy to draw. Prediction is the whole game.

Here's what Orivra does for prediction and typing:

  • Word suggestions and next-word predictions that chain β€” keep tapping and you can type out whole sentences, not just the next word
  • No-space rescue: I type whole phrases as one stream and it works out the word breaks when I commit
  • Fat-finger correction that catches my real errors, and literal protection β€” type a word exactly and hit Space, it stays
  • Swipe / glide typing, built for sloppy swiping, with a Tap / Swipe mode switch so it never breaks my tap flow
  • Prediction strip swipes, one of my most-used features: swipe down on a chip to take it, left and right to move, up for alternatives
  • Rise: swipe up and the strip opens into an expanded panel above the bar with more suggestions in view β€” including long sentence predictions shown as one full-width row, so a whole sentence I actually say can be offered and taken in one tap instead of being chopped into little pieces
  • Single-canvas layout, built and installed on Nov 12, kept alongside the classic layout as a fallback

Under that is a big local vocabulary: a dictionary of tens of thousands of words, a large ranked everyday pack, and next-word tables built from how people actually talk. Slang, brand names, and edgy words are in on purpose, because that's how I type.

It learns how I type β€” and I can see everything it learned

Orivra learns correction pairs three ways: when I pick a correction, when I delete and retype, and when I highlight a mistake β€” even a whole multi-word stream β€” and type the fix over it.

Fixes I make by highlighting wait for approval first. Nothing goes live until the same fix repeats or I approve it myself in the Saved Lab, where I can see, approve, or delete every single thing it has learned.

There's also Manual Teach, a personal dictionary, and a personal next-word model that survives restarts. Different ways of teaching it are kept separate, so a casual tap doesn't teach it the same thing as a real correction.

And my own data doesn't win on one random tap. My personal predictions only take over after I've actually done something repeatedly. One-offs don't get to hijack my suggestions.

Later it got literal. We shipped a typing heatmap in build 4353, built from touch data Orivra already stored. First readout: 26,372 taps, my P, L, and R are off-center, C, T, and K are scattered, and I'm 97% calibrated overall. I was digging it right away, because it wasn't guessing about me β€” it was showing me.

Spanish wasn't an afterthought

English came first. Then I built Spanish properly β€” northern Mexican Spanish and border Spanglish, how I actually talk.

The full Spanish pack was built and verified on Nov 8. It was functionally complete and installed on Nov 13, with its own prediction and learning stores, completely separate from English. Γ‘ key, long-press accents, Β‘ and ΒΏ.

Themes, because I look at this thing all day

Orivra has 10 themes, light and dark.

My favorite is the app-adaptive one. Orivra derives its Material colors from the app I'm in β€” Music gives it green, Messages blue, Mail red, Notes amber, Photos violet. The board, the keys, the highlighted prediction chip, and Enter all change together. I can override it with a manual color, and go back to app color in one tap.

Settings got a full redesign in build 4351 too β€” Spanish counts visible, a proper Manage door, the Spanish Saved Lab.

The rest of my daily kit is in there too: emoji browser with search, skin tones, and recents, symbol layers and long-press characters on the keys, split-keyboard mode, adjustable long-press speed, cursor magnifier, and backup and export of everything it learned β€” as my file, on my phone.

One thing I won't pretend is built: clipboard history. It's planned and designed, and queued. Not shipped yet.

Nothing leaves my phone

Orivra is fully offline. It doesn't even have internet permission. No account, no cloud, no telemetry.

It doesn't learn from password fields. Our measurement work is counts only β€” no text stored, nothing sent anywhere.

A keyboard sees your passwords and your banking. That privacy stuff can't be something you check at the end. It has to be in every spec, from the start.

Where it is now

As of today Orivra is on build 4354, installed and live, with 4355 in the works.

The work ahead is the Learning Smartness list: my personal predictions outranking the generic pack, a window where I can inspect and edit what it learned about my next words, automatic phrase learning, and later a Spanish personal learner with a Spanglish bridge β€” with English and Spanish stores still kept separate.

Not done β€” and that's fine

People ask me every day if it's done.

No. A keyboard that learns how I type probably never really gets to be done. It just keeps living my real day with me β€” 13-hour shifts, heat, battery, switching languages, typing fast and badly β€” and keeps getting better at typing like me.

My favorite part is the full circle. I started as a swipe typer on Gboard. Building Orivra sent me back to the stone ages, tapping everything, while we built the prediction good enough to live on. Then swipe came home, on a keyboard I built myself. It's been a fun learning thing.

I'm not posting this to sell anything, and there's nothing to download. I just want to talk about it.

So here's my question, the same one I built Orivra around: if your keyboard was built for how you actually type β€” not how you're supposed to type β€” what would it do different?

Tell me. That's the conversation I want to have.

β€” Javs

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.