You find out what you don’t understand when you try to explain it.
Rereading slides feels like studying. It mostly tests whether the slides look familiar. Explaining a concept out loud tests whether you understand it, but alone you can’t tell which parts of your explanation were wrong.
Kizuki is a study tool that runs on your computer. You add course material, pick a concept, and explain it without notes. Kizuki plays the student. It asks about what you got wrong, what you left out, and what you said vaguely, and every question quotes your own material.
The model points. It never writes the facts.
A small model is good at noticing that two passages might disagree. It is bad at stating facts. So Kizuki only lets the model point at sentences in your material, and the app shows those sentences word for word, each with a link to the page, slide, or cell it came from. If the model can’t point at a sentence, it can’t make a claim.
Your material can be wrong
When your explanation and the material disagree, Kizuki shows the passage and asks which is right: the material, you, or neither. A slide can have a typo, or a professor can correct it in class. Once you say the material is wrong, your correction wins from then on.
Nothing is saved without your OK
Proposed concepts, the order they build on each other, and the list of what you missed all wait for you to confirm them. At the end of a session you tick the sentences you did miss, so the review schedule runs on your judgment, not the model’s guess.
Reviews follow what you missed
A clean session doubles the wait before the next review: 1, 2, 4, 8 days, up to 60. A session with misses brings the concept back tomorrow. A concept waits until the ones it depends on are solid, and setting an exam date pulls reviews earlier.
What I owned
I built the local server and dashboard, the readers for PDF, PowerPoint, Word, Excel, CSV, and Markdown files, the review scheduler, and the storage. I also built the model tests: they plant mistakes in an explanation and count how many Kizuki catches, so different models can be compared on the same test.
Local by default
Kizuki runs a small model through Ollama. There is no account and no cloud. Everything lives in one folder: copies of your files, append-only event logs that are the record of truth, and a SQLite search index that Kizuki can rebuild from the logs at any time. To back up your study data, you copy the folder.
You can point Kizuki at a hosted model that speaks the OpenAI format. Before that would send your material off your computer, Settings asks you to confirm it.
Where it stands
Kizuki is open source under Apache-2.0 and on npm. npx kizuki checks for the two Ollama models, starts the app, and opens the dashboard. The docs are generated from the code, and the build fails when an export or a route has no docs.