The problem
I enjoy learning languages, but I was dissatisfied with the apps available to me. Most either prescribed the material, reduced practice to a short daily game, or covered one skill without giving me a place to connect it to the others. That did not match how I study Chinese, especially at the intermediate stage, where reading real material becomes the most important thing you can do and also the most frustrating, because you stop every few words to switch to a dictionary.
I built Lexiwyn for my own study first. It is for adult learners of Mandarin who know somewhere between a few hundred and a few thousand words, roughly HSK 2 to 5, and who have outgrown textbook dialogues and gamified apps. It is meant for a deliberate study session, not something to tap through for five minutes on the train.
What I built
Lexiwyn is built around one idea: reading is the engine. A learner reads a graded text from the library or imports their own material from an article URL, a YouTube link, pasted text, a PDF, or a video. Each text shows what share of its words the learner already knows. Tapping any word gives its pinyin and meaning, looked up in a local CC-CEDICT dictionary first and only falling back to Gemini when that is not enough.
Every word the learner looks up goes into a single word bank, and that word bank drives everything else. Words come back in spaced-repetition reviews, get used in AI tutor lessons, and turn up in conversation practice, and each of those raises the same per-word familiarity score that the reader shows as a highlight. Other apps offer these tools separately. The connected loop is the point. AI explains, generates practice, checks writing, and holds conversations, but it supports the learner's effort instead of doing the learning for them.
Graded texts or anything you want to read
The library is where a study session starts. Pick a graded story or import your own, and each one shows how much of it you already know.
- Import from an article URL, YouTube link, pasted text, PDF, or video
- Each text shows the share of words you already know
- Graded stories are tagged by HSK level
Decisions and tradeoffs
One word bank, not separate tools
Reviews, tutor lessons, and conversation all read from and write to the same per-word familiarity score. That is harder to build than three independent features, but it is what lets a word tapped in a story come back in a review and then in a lesson.
Dictionary first, LLM last
Word lookups go to a local CC-CEDICT dictionary, then a cache, and only then to Gemini. Most taps are answered instantly and at no cost, and the model is kept for the cases where context actually matters.
FSRS instead of an LLM scheduler
Review scheduling uses the FSRS algorithm rather than asking an LLM what to review. It is deterministic and predictable, leaving the AI for tasks where generated language is useful: explanations, conversation, exercises, and feedback.
Cost observability from day one
Every Gemini call is metered by user and feature. Knowing what each part of the product costs is what made it possible to set free and Pro limits that leave reading and reviews uncapped.
What I learned
Building the features was only part of the work. Using Lexiwyn during my own study showed me that small interruptions matter: an extra step before reading, or a page covered in underlines on a hard text, can break a focused session. Several of the reader's defaults changed because of that.
It also made the right role for AI clearer. Generated language is useful for explanations, feedback, and conversation, but lookups, progress tracking, and review scheduling are easier to trust when they stay deterministic.
Result
Lexiwyn is live at lexiwyn.com. The full loop works across reading, word lookup, reviews, tutor lessons, writing feedback, grammar practice, video study, exam mode, and voice conversation, with a placement test at onboarding and HSK 3.0 level progress. Reading, lookups, the word bank, and reviews are free without caps. A Pro plan, billed through Stripe, raises the limits on the AI-heavy features such as tutor lessons and live voice conversation.
Alongside the rename from Blueprint, I redesigned the whole interface around a printed picture-book look and moved hosting to Cloud Run and Cloudflare. I still use it for my own Chinese study, which is how I find the small problems that only show up in real, repeated use.
Since the rename the work has mostly followed what I wanted while studying. A whole PDF book can now be read inside the app, either as clean text or as the original pages, with chapters found from the printed headings. Anki decks can be imported with their review history, so years of flashcards do not have to start from zero, and the same word in two decks now shares one review schedule. There is also a public word of the day, and Review, Exam, and the tutor pages have been redesigned to use more of the screen.