Fillr

Inspiration

The idea for Fillr came from a problem we have both had in lectures: sometimes professors simply talk too fast. You either spend the entire class trying to write everything down and stop properly listening, or you focus on understanding what is being said and end up with incomplete notes afterward.

This led us to guided notes, and there is actually a good amount of education research behind them. Studies have found that partially completed and structured notes can help students stay engaged, reduce cognitive load, and in some cases improve learning compared with simply giving students complete notes. The catch is that the design matters a lot: if you provide too much information, students become passive, but if you provide too little, you have basically just given them a blank page.

The other problem is that making good guided notes takes a lot of work. Instructors have to go through every lecture, decide what structure to provide, what information to leave out, and where students should actually have to think. Even if professors know guided notes can be useful, doing that manually for every lecture is not very realistic. That is where we thought AI could actually help: not by writing the student's notes for them, but by doing the tedious work of creating the scaffold so the student can focus on learning.

What It Does

Fillr takes lecture slides or notes and turns them into a guided note template that students can use during class. It keeps the structure of the lecture, including the main topics, sections, and useful examples, while leaving enough open space for the student to write down the actual content themselves.

For example, instead of giving a student a full explanation of stacks, Fillr might generate:

Stacks

[space for student to take notes]

Example: a stack of plates

The goal is not to generate finished notes for the student. It is to give them a better starting point for taking their own notes.

How We Built It

Our stack:

Layer	Technology
AI Model	NVIDIA Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8
Model Hosting	NVIDIA Brev
Frontend	React
Backend	Python
API	FastAPI
Database	PostgreSQL

The user uploads their lecture material, which is passed to our model. The model analyzes the structure and content of the lecture and returns a guided-note template.

One of our biggest technical changes was switching the model output from LaTeX to JSON. Originally, we asked the model to generate the final template directly in LaTeX, but it was inconsistent and would often break the syntax. JSON was much more reliable, so we now have the model generate structured JSON and let our own code handle the final formatting.

Challenges

Finding the right amount of information

This was probably the hardest part of the project. Our early templates usually went to one of two extremes: sometimes Fillr basically copied the slides into the template and left very little for the student to do, while other times we removed so much information that the template was just a lecture title and a few section headings.

The research we looked at helped us think more carefully about where that balance should be. Guided notes should provide enough structure that students are not spending the entire lecture figuring out how to organize the page, but they should still require the student to actually listen, understand, and write down the important ideas themselves. We also found research suggesting that more open-ended outline notes can lead to better recall and inference than simple fill-in-the-blank notes, so we tried to make Fillr create useful structure rather than just delete random words from sentences.

The model kept giving away too much

LLMs are very good at summarizing things, which turned out to be a problem for us. If we gave the model a set of lecture slides, its natural instinct was to create a clean summary of them, but that completely defeats the purpose of Fillr. We had to prompt it to first understand which information is important, and then separately decide which parts should actually be shown to the student and which parts should be left out.

Prompt instructions leaking into the notes

We also had an issue where the model started including our own instructions inside the generated template. For example, we would tell it things like "do not make this an answer key" or "leave important ideas for the student to fill in," and then those instructions would appear in the actual notes. We had to make the output format stricter and filter out anything that was meant for the model rather than the user.

What We Learned

The biggest thing we learned is that building Fillr is mostly about deciding what the AI should not give you. We also learned that JSON is much more reliable than asking an LLM to directly generate complicated LaTeX, and that a good guided note is not just a normal set of notes with random words removed. Too much scaffolding turns the template into an answer key, while too little makes it useless, so most of the work came down to finding a useful middle ground.

More broadly, we liked the idea that AI does not always have to make something easier by doing it for you. In this case, it can take care of the tedious work of structuring the notes while still leaving the actual thinking to the student.

What's Next

We want to keep improving how Fillr decides how much information to include, especially because different subjects probably need very different styles of templates. We also want to add different template styles, better export options, adjustable levels of guidance, and an after-class mode that turns a student's completed notes into flashcards and practice questions.

Long term, we want Fillr to stay focused on the same idea we started with: using AI to help students learn without replacing the process of learning itself.