Sayvero

Inspiration

More than a billion people speak English as a second language, and most of them use it every day: at work, at the store, on the phone. Many still make the same small mistakes they made years ago, because nobody corrects them. People understand them, so why would they? Apps teach scripted sentences and a tutor hears you an hour a week, but the mistakes that stick happen in the rest of the week. Sayvero started from a simple question: what if a coach could learn from the way people actually talk?

What it does

Sayvero is an English coach that listens through your Bee. It picks out the mistakes you really make in conversation and groups them into patterns: "explain me the difference" and "explain you the plan" count as one habit, explain me → explain to me. The app shows your week as a short list, sorted by how often each habit happens. Every card shows the sentence you actually said, the part that's off, and how a native speaker would say it.

How I built it

• Bee CLI: bee proxy exposes Bee's live stream locally. Sayvero subscribes to its new-utterance server-sent events, so each sentence arrives seconds after you say it.
• Agentic AI pipeline: each sentence goes through an agentic pipeline that separates real learner mistakes from normal speech, explains them in plain language, and builds a picture of your habits over time.
• Backend and app: a Python API and a mobile-first React Native app.

Challenges

• Speech isn't writing. Real conversation is full of restarts, fragments and filler. Sayvero has to leave all of that alone and only flag what a native speaker wouldn't say, while ignoring anything that could be a speech-to-text error.
• The same mistake, said a hundred ways. A habit doesn't repeat word for word, so recognizing that two different sentences hide the same mistake took a lot of iteration.
• The world is noisy. Bee hears everything, including videos and voice demos playing nearby. Telling a learner's speech apart from everything else around them is an ongoing challenge.
• Real time without waste. Reacting within seconds of a sentence, without spending compute when there's nothing new to do.

What I learned

The hard part of a language coach isn't finding mistakes. It's not flagging the thousand things that only look like mistakes in a transcript. Precision matters more than recall: one bad correction costs more trust than one missed mistake.

What's next

Practice mode for each habit, progress over weeks, and a more resilient live pipeline and language coaches for more learners of other languages like Spanish.

Built With

• agentic-ai
• bee
• bee-cli
• expo.io
• fastapi
• open-ai
• python
• react-native
• typescript