Sloth Study Digest: Lecture Summaries and Flashcards While I Sleep, Fully Local

rudra dixit — Fri, 23 Oct 2026 02:20:27 +0000

These screenshots work for the post. The flashcards you flipped show their answers, and the third card still shows "Tap to reveal the answer", so the page does what it should.

On quality, the recall of topics is good and RMMM (Risk Mitigation, Monitoring, and Management) matches the usual textbook term. Two weaknesses are worth admitting in the post. The "Key terms" are mostly acronym expansions (LOC: Lines of Code) rather than real definitions. The flashcard "What are the steps in a software process?" with the answer "Documentation, implementation, deployment, interaction" looks off to me. I can't check it without your PDF, so look it up in the chapter. If it's wrong, that's a useful concrete example for the limitations line.

Here's a full draft with your real numbers. I've assumed you only tested PDFs, so it says audio support is built but untested. Replace the bracketed parts.

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady

What I Built

Sloth Study Digest is a tool for students with a pile of reading and no energy to summarise it. You drop PDFs, notes, or lecture recordings into an inbox/ folder and run one command. A few minutes later you have a summary page for each file with key points, key terms, and a set of flashcards you can click to flip.

It's for anyone with a study backlog. The slow, tedious part (reading and note-making) happens while you do something else.

Demo

[Screenshot: the index page, "Good morning. Here is what got done overnight."]
[Screenshot: one digest page with summary and key points]
[Screenshot: a flashcard with its answer revealed]

Code

ITZMElk / sloth-study-digest

Sloth Study Digest

Drop lecture recordings, PDFs and notes into inbox/, go to sleep, wake up to a summary page and flashcards for each one. All local, all open-source:

audio -> whisper.cpp -> transcript -> Ollama (open-weight model) -> HTML digest

No API keys, no uploads, no internet needed after setup. Python standard library only (pypdf is optional, for PDFs).

Setup (once)

Install Ollama and pull a small model ollama pull qwen2.5:3b (or llama3.2:3b)
For audio/video: install ffmpeg, build whisper.cpp, and download a model, e.g. ggml-base.en.bin.
For PDFs: pip install -r requirements.txt

Use

python sloth_digest.py                       # process everything new in inbox/, then exit
python sloth_digest.py --watch 300           # keep running, check every 5 minutes
python sloth_digest.py --model llama3.2:3b   # swap the model, nothing else changes
python sloth_digest.py --whisper-bin C:\whisper\whisper-cli.exe --whisper-model C:\whisper\ggml-base.en.bin

Open digests/index.html in the morning. Click a flashcard to reveal its answer Files…

View on GitHub

How I Built It

Model: qwen2.5:3b, a small open-weight model, run locally with Ollama.
Audio (built, not yet tested): whisper.cpp is wired in to transcribe lecture recordings. I only tested PDFs this week.
Glue: one Python file using only the standard library, plus pypdf for PDFs.

The pipeline is deliberately simple. It extracts the text and cuts it into chunks. The model writes notes for each chunk, then turns the notes into a three-sentence summary, key points, and key terms, and finally into flashcards as JSON. Everything is written out as plain HTML with no CDN or web fonts, so the results open offline. Files already processed are skipped, and a file that fails is logged without stopping the rest.

Why Does Open Innovation Matter?

My course material stays on my laptop. Nothing is uploaded to a server I don't control.
It costs nothing to run. There are no API keys, per-token bills, or rate limits, so I can run it on every chapter without thinking about price.
It works offline after setup. The model runs locally, and the output pages need no internet.
I can change the model. It's one command-line flag (--model), and I can edit the prompts freely. A closed API wouldn't let me tune any of that.
Slow is fine. A 3B model on a normal laptop is too slow for live use, but for "run it and walk away" it's plenty.

The Lazy Payoff

I ran it on three chapters on software project management, about 8,000 words in total:

File	Words	Time	Flashcards
Chapter 4	4,377	60 s	8
Chapter 5	1,604	27 s	8
Chapter 6	1,981	38 s	8

That's about 2 minutes 10 seconds in total, with the model using 2.16 GB of memory, on [your laptop specs]. Doing this by hand would take me [your honest estimate].

Limitations. A 3B model is good at picking out the topics but shallow on detail. Key terms are often just acronym expansions rather than definitions, and I found [at least one flashcard answer that didn't match the chapter: describe it, or delete this sentence if all your checked cards were right]. I treat the digest as a quick outline before reading, not a replacement for the chapter.

Before you post:

Check that flashcard against the chapter and fill in or delete the limitations sentence.
Add your laptop specs (RAM and CPU) and your own time estimate.
Delete the "Prize Categories" and "My Agent Session" sections unless you have a partner category or a DevRelay session link.
If you want to claim audio, run one short recording first and I'll update the "How I Built It" section.

DEV Community: rudra dixit

[Boost]