Kriah Listener

I'm Rabbi Dr. Levi Druin, principal of a Jewish day school, and I've never studied coding. From their first days in elementary school, our children learn kriah, which means reading Hebrew with its vowels. A teacher can't hear every child read every day, and the mistakes that matter are tiny: one vowel under one letter, or one dot that turns a b into a v. Children also read in different traditions. In Ashkenazi reading שַׁבָּת is sha-bos, and in Sephardic reading it's sha-bat. A speech tool that expects one tradition marks the other one wrong. So I built Kriah Listener, a station that listens. A child taps their animal and reads, line after line, with the microphone on. Muse Wonder hears the sounds, and it never sees the text, because a model that knows the word tends to hear the word. Meta's open phoneme model checks the vowels against the child's own voice. Then plain code marks every letter and vowel by the child's own tradition. Green means right, orange a wrong letter, purple a wrong vowel. What the child said is crossed out, with the right sound under it. The next lines get harder or easier depending on how the child read. The teacher gets a reading map for each child, with every letter and every vowel marked solid, shaky, or not tested yet, plus practice groups and a printable packet. Children can practice every day, not only when the teacher gets to them. The teacher sees which letter and vowel each child is missing without listening to every line. I tested it on my own voice, with mistakes planted on purpose. No child has used it yet. The audio goes to the Meta Model API, and I won't send a child's voice there until a parent signs a consent form a lawyer has read. The station never saves audio to disk or to a log. Before every test round, I locked the pass rules so I couldn't fool myself. The first five rounds failed, and I published every one with my fixes. The sixth passed in all three runs. It caught all 4 planted mistakes every time and didn't mark a single correct reading wrong. Muse Wonder alone made 5 or 6 false marks a run. Live, a line grades in about 3 to 5 seconds. Next comes the consent form, then a round with children, a check of the rulebook by our kriah teachers, and a pilot in one first-grade classroom. With a new rulebook, it could listen in any language. Every child deserves to be heard. I built it all during the event, starting with my project brief on October 26 (commit 3fbef66). Claude Code and OpenAI Codex wrote the code from my brief, and Codex and Muse Code reviewed it. I designed it, recorded every test reading, set every pass rule, and decided what counts as a mistake. Try it: https://levis-new-mac-mini.tail8bd372.ts.net (Chrome, or Safari on iPhone). Tap Guest to read Sephardi, after a short voice check. Demo teacher view, no PIN: https://levis-new-mac-mini.tail8bd372.ts.net/?mock#/teacher