SlothRead：懶得讀 40 頁論文？讓本地 AI 為你製作管理層簡報

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

What I Built

I built SlothRead, a privacy-first, offline executive synthesizer for EPUB books and dense research PDFs.

Between 40-page arXiv whitepapers, technical documentation, and long-form e-books, reading backlogs constantly pile up. In the spirit of International Sloth Day, SlothRead rewards digital laziness by automating the cognitive heavy lifting.

Instead of forcing you to skim dry prose and suffer cognitive burnout, you simply drop a PDF or EPUB into SlothRead. Running a local open-weight model completely on your machine, it distills the entire piece into a sleek, dark-mode Executive Presentation Deck featuring:

3 Non-Negotiable Core Takeaways per chapter.
An Interactive Technical Glossary explaining domain jargon upon hover.
A 60-Second Audio Recap generated completely offline using your browser's native Web Speech synthesis.
A Global Executive Brief detailing the core thesis and the exact reading time saved.

Researchers, developers, and students can conserve their energy like a sloth and digest complex documents in under 2 minutes.

Demo

Experience the live app or inspect the hosted instances directly:

Frontend Interface (Vercel): https://slot-read-ewx9.vercel.app

Code

Repository Link: https://github.com/ziyadparkar01/SlotRead

Repository Architecture

/backend: FastAPI service handling section/chapter chunking with PyPDF and EbookLib, alongside async Ollama inference pipelines.
/frontend: Decoupled vanilla HTML5/ES6/CSS interface featuring glassmorphic editorial cards, responsive chapter navigation, and a reactive 3D canvas background built with Three.js.

How I Built It

SlothRead is built entirely around an open-source, local-first inference architecture:

1. Local Open-Weight Inference Engine

Inference runs locally via Ollama using Google's Gemma 2 (gemma2:2b) or Qwen 2.5 (qwen2.5:3b).
Configured with native JSON mode (format: "json") paired with strict schema enforcement to guarantee consistent, structured outputs across chapter summaries, glossary terms, and narration scripts.

2. Document Parsing & Chunking Engine (`backend/app/parser.py`)

Built on pypdf, ebooklib, and beautifulsoup4.
Analyzes headings, chapter markers, and document hierarchies to extract coherent, meaningful narrative sections rather than relying on naive token splits that cut sentences mid-thought.

3. Decoupled Modern Frontend

Pure Vanilla ES6+ JavaScript — zero heavy framework dependencies or TypeScript compilation overhead.
Designed with an editorial aesthetic using Cabinet Grotesk, Newsreader serif typography, and JetBrains Mono.
Features an interactive 3D particle background rendered using Three.js that responds smoothly to mouse parallax and scroll position.

Why Does Open Innovation Matter?

Open innovation is not just an optional detail for SlothRead; it is the fundamental reason the project can exist safely:

Uncompromised Document & IP Privacy: Research papers often contain unreleased patents, proprietary company methodologies, or confidential legal and financial documents. Uploading these files to closed, centralized cloud APIs introduces serious data leakage risks. By running local open weights, confidential files never leave the host machine.
$0 Infinite Inference: Passing a 300-page book through commercial cloud endpoints racks up heavy token bills and frequently crashes against API context caps. Open-weight models running on consumer hardware cost absolutely nothing to run.
True Offline Independence: Whether you are commuting on a train, sitting on an airplane with no Wi-Fi, or working behind an air-gapped corporate firewall, SlothRead operates 100% offline without requiring an external internet connection.

The Lazy Payoff

During testing, I fed SlothRead a dense 38-page research whitepaper on distributed consensus mechanisms.

Traditional reading time: ~45–60 minutes of manual reading, skimming, and note-taking.
The SlothRead workflow: 1 drag-and-drop, ~40 seconds of background local processing on my laptop, and a crisp 4-slide executive presentation deck.
Total effort: 2 clicks. I stepped away, hit "60s Listen" on the executive summary, and walked away with every critical takeaway and term defined. *