Scrolling past paragraphs of filler text just to find recipe steps or DIY instructions? JustTheInstructions is a Google Chrome Extension which solves this issue by detecting instructional content in real time as soon as a page loads — delivering clean, formatted instructions in just one click.

• Automatic Detection: Pages are analyzed locally in under 500ms using a custom ONNX model.
• Smart Notifications: A Chrome notification instantly appears when instructional content is detected — no need to open the extension.
• One-Click Extraction: Click once, and the instructions are automatically extracted, cleaned, and formatted for you — ready in seconds instead of endless scrolling.
• Optional Manual Mode: Notifications can be toggled off — you can still run the extension manually in just 2 clicks (open extension → click “Extract Instructions”).

• Real-Time Detection: Page content scored for instructional relevance in 300–500ms after load
• One-Click Flow: Notification → click → clean instructions
• Model Accuracy: 99.6% on a 100k-sentence test set (80/10/10 split)
• Inference Speed: ~11ms/sentence (~80+ sentences/sec)
• Extraction Speed: Formatted instructions returned in ~2–10 seconds
• Model Size: 2.5MB ONNX — small enough for fast extension loading & smooth performance on low-end laptops
• Data Scale: ~1M labeled sentences scraped from 5,000 instructional websites
• Privacy: Detection is 100% client-side; formatting only runs when you explicitly request it

1. Score instructional relevance with >99% accuracy while running entirely client-side.
2. Deliver clean, formatted instructions in under 10 seconds end-to-end.
3. Eliminate manual interaction — automated detection + one-click flow, with manual fallback.
4. Maintain privacy-first design by limiting server calls to formatting only.

1. Real-Time Detection (Client-Side)

   • As soon as a page loads, text is segmented into sentences.
   • A TensorFlow-trained binary classifier (exported to ONNX) assigns each page a confidence score (low/medium/high instructional content).
   • Detection runs fully in-browser via ONNX Runtime (WebAssembly).

2. Why ONNX Runtime?

   • 2.5MB model size — ~40% smaller than a comparable TensorFlow.js build, reducing download & memory usage
   • Faster inference due to WebAssembly optimizations
   • Better for weak devices: Lightweight enough for low-end laptops and Chromebooks

3. Notification-Driven UX

   • If the score passes a threshold, you receive a Chrome notification — no need to open the extension.
   • Clicking the notification triggers automatic extraction.
   • Notifications Off? You can still run the model manually in 2 clicks (open → extract).

4. Backend-Powered Formatting

   • Candidate sentences are sent to a lightweight Flask API, which:
     • Cleans & structures the text into Markdown-formatted steps
     • Adds prerequisites (ingredients, tools, etc.) when relevant
   • The API is Dockerized and deployed on GCP Cloud Run for scalable, low-latency performance.

5. Privacy & Safety

   • Detection always runs locally — only when you explicitly click “Extract Instructions” is page data sent for formatting.
   • Sensitive domains (e.g., Gmail, banking) are automatically skipped.

• Scraped 5,000 instructional websites across five categories:
  🍲 Recipes • 🎨 Crafts • 🔌 Circuits • 🧪 DIY Science • 🛠️ General Tutorials
• Collected ~1M labeled sentences via a custom BeautifulSoup + Pandas pipeline.
• Dataset split into 80% training / 10% test / 10% validation.
• Achieved 99.6% test accuracy (100k sentences).
• Model exported to ONNX Runtime for optimized frontend inference.

Model Architecture:
Binary text classifier — Embedding → GlobalAveragePooling → Dense(16, ReLU) → Dropout(0.4) → Dense(16, ReLU) → Dropout(0.4) → Sigmoid (~64K parameters, Adam optimizer)

📄 Colab Notebooks (for reproducibility):

• Model Architecture & Training
• Data Collection & Scraping Pipeline

⭐️ Found this useful? — Star this repo
📝 Help others discover it — Leave a review on the Chrome Web Store