- Live: hayashi.in/quant
I wanted to learn quantitative asset management the way people learn a language: a few minutes a day, lots of repetition, and each idea seen from several sides until it sticks. Flashcards didn’t fit, because recognizing an answer isn’t the same as being able to explain it. So in Quant Fluency you write your answer in a sentence or two, and a model grades how well you understood the idea.
The course map has 227 concepts, from what a bond is up to Black-Litterman, risk parity and backtest overfitting. I built it from the topics covered by the CFA, FRM, CQF, CAIA and CIPM programs (the site isn’t affiliated with any of them) and arranged it in eight tiers with prerequisites, so a concept unlocks once you’ve seen the ones it builds on. The first tier is written: 21 concepts, each with a lesson, a longer “More detail” section with worked examples and links to primary sources such as the SEC, FINRA and CFA Institute, and three questions that come at it from different angles: explain it, apply it, calculate it, spot the misconception.
How it works
- On your first visit you rate how familiar you are with each of the 16 areas, from 1 (new to it) to 5 (could teach it). Every concept has a level within its area, so the app skips what you say you know and starts at your level. Miss two of your last three questions in an area and it steps down a level. You can also change the levels yourself at any time.
- You answer in writing. There’s no multiple choice. A qwen3.5 9B model running on my home server through Ollama compares your answer with a reference answer and a list of key points, scores it from 0 to 100, and tells you what you left out. It grades meaning, not wording.
- Two numbers sit at the top of the page. Mastered is the share of concepts you’ve been shown that you now know. Covered is the share of the whole course you’ve been shown at all. The first says how well you know what you’ve seen. The second says how much you haven’t seen yet.
- Passing a concept schedules it to come back after 1, 3, 7, 21 and then 60 days, each time as a different question. A concept counts as mastered after three passes spread over at least four days, on at least two different questions, so nothing is mastered on the first day.
- After each answer you can review the lesson, follow the references to check it, or ask a tutor follow-up questions, up to five per card. Diagrams, in the lessons and in the tutor’s answers, are drawn with Mermaid.
- Progress lives in your browser. Signing in is optional, with Google, GitHub or an emailed link. If you do, your progress and levels follow you between devices, and my server keeps your tutor conversations and an access log for 180 days. You can download or delete all of it from the app. The text of graded answers isn’t stored either way.
What I ran into
A small model grades well but explains badly. Grading against a reference answer is a narrow job, and the 9B model does it well in about two seconds: in my tests it gave zero to wrong answers and to answers that tried to talk it into a perfect score. Explaining is harder. When I ran 27 realistic follow-up questions through the same model, more than a third of its explanations had a real error, like a backwards bid and ask or an arithmetic slip, and the charts it drew were often wrong. A stricter prompt helped but didn’t fix it, and turning on its reasoning mode made things worse: it spent its whole budget thinking and returned nothing. So grading stays on the home server, and the tutor uses a larger model through Ollama’s cloud service, which got every number right when I reran the questions the small model missed. That means your follow-up questions leave my server, while your graded answers don’t.
The browser never sends the right answer. It would be simpler for the page to send your answer and the reference answer together, but then the endpoint would grade anything anyone sent it, which makes it a free chatbot on my GPU. The page sends only a question id and your text, and the server looks up the reference itself. The tutor works the same way, with a limit on questions per card and an instruction to stay on the course’s topics.
Charts need to survive a phone screen. Mermaid only draws a chart legend when the chart is about 600 pixels wide, so on a phone a two-line chart showed two colors and no labels. The page now draws charts at the width they’re shown and adds its own legend underneath.
The next tier covers the statistics underneath everything else: variance, correlation, regression, and the distributions that show up in risk models.