Chezky Kohn
Search that understands what you're actually looking for. Scheduling that gets a chavrusa on the calendar in two clicks. Speech recognition for Yiddish, which no one else was building. Tools for the long, patient work of putting seforim together.
Most of it started because I wanted the thing to exist and it didn't. Some of it is client work — computer vision on a production floor, a dictation app for someone who works in Yiddish all day. I'm in Brooklyn, usually running several of these at once, and always glad to talk to someone working on something adjacent.
Fifteen minutes of learning with one person at a time. You send someone a link, they pick a slot from your real availability and choose a sugya by searching Sefaria, and it lands on both calendars with a Meet link. Built around a simple idea: the hard part of learning with people isn't the learning, it's the scheduling.
Booking links and group polls that check my actual calendar before offering a time, so nothing gets offered that I can't take. Requests reach me by email the moment they're made. Same Google connection as Chavrusa 50 — one calendar, two front doors.
A search and indexing layer over a large Torah corpus — semantic and keyword search across books, commentaries, and passages, with a hand-built taxonomy of entities and an index that grows as it's used. For the moment you half-remember something and need to find where it actually says it.
Speech recognition for Yiddish and Hebrew. Off-the-shelf models handle them badly, so this is training and evaluation from the ground up: audio collection, transcript alignment and review, fine-tuning runs, and honest benchmarking of what's actually improving rather than what looks good in a demo.
Structuring Torah material into lessons that build on each other — chapters, sequencing, sources, and versions, with the pedagogy written down instead of living in someone's head.
A food production facility wanted per-worker output counts without clipboards. Four side-view cameras, YOLOv11 detections projected from pixels into a shared millimetre coordinate plane on the table, clustered across feeds so the same object isn't counted twice, then assigned to worker zones. Out comes annotated video and a CSV at the end of the shift.
Windows has no Yiddish dictation. Neither does anything else. This is a tray app with a global hotkey: press it anywhere — Word, email, a chat window — speak, and the transcription types itself in wherever your cursor was. Built on a custom Whisper deployment for a client who works in Yiddish audio all day.
Duolingo doesn't have Chumash, and Hebrew apps assume a Modern Hebrew learner. This one is built for Yiddish-speaking kids learning the vocabulary of Parshas Vayikra — and built ADHD-first, because in that classroom it isn't an edge case. No timers, ever. Large touch targets. Corrections without shame.
Twenty-four volumes spanning על התורה, מועדים, תהילים, כל הש"ס, משלי, אבות and אגדות.
A way of building a maamar that I keep coming back to: six מעיינות — הראשונים, מדרש וזוהר, הדרוש, החידוש, המוסר, הסיפור — so a piece moves from the sources through the chiddush and out to something you can actually say over.
A searchable corpus of the seforim above alongside לב אהרן, אוצרותיהם של צדיקים, and roughly three hundred weekly gilyonos. Increasingly the thing the software above points at.
I write up most of this as I go — how the pieces actually work, what broke, and where Yiddish speech recognition stands right now.
The fastest way to reach me is to grab a time — you'll only see slots I can genuinely make.