Semantix

The problem

Language apps make you tap through drills that were written before you showed up. You memorize a deck and still freeze in a real conversation - because the practice and the review were built as two separate products, and neither one knows what you personally keep getting wrong.

Our approach

We built the review out of the conversation instead of beside it. Learners talk to a tutor running on Claude via AWS Bedrock, immersed in the target language with light English scaffolding. What the learner fumbles becomes an Anki-style card scheduled by SM-2, so the deck is a record of their own mistakes rather than a generic word list. Around that sit practice modes, writing correction, and a hand-authored plus AI-generated curriculum.

Cards generated from real conversations

The tutor writes flashcards from what the learner actually got wrong, scheduled with SM-2 spaced repetition. No shared deck, no manual card entry.

19 languages, A1 to C1

A complete structured path per language, combining hand-authored curriculum with generated material, plus placement testing so learners start at the right level.

Six practice modes beyond chat

Cloze, translation, conjugation, scenario, pronunciation, and dictation - with Polly neural voices for speech and native iOS speech recognition in the app.

Single-table DynamoDB, tested against the real engine

Local development and the test suite run against DynamoDB Local rather than a stand-in store, so the tests exercise the same engine semantics production does.

Shipped to iOS without a second codebase

A Capacitor shell around the live site plus three native Swift plugins for speech, StoreKit 2, and haptics - one product, not a web app and a separate rewrite.