[Oh_Well]

My journey building a chrome extension for
doomscrolling awareness + Facebook post analysis

Hey guys, as my 1st entry to this awesome weekly contest by Quantum

https://hackforums.net/showthread.php?tid=6310391

I wanted to share a personal project this time, that I started recently.

A period passed when I wasn't good, I had dark thoughts and the algorithms weren't helping. I found myself scrolling endlessly on negative-vibed feeds for hours, until I decided that I didn't want to do that anymore.

So I build a Chrome extension that tackles doomscrolling while also experimenting with Facebook post analysis.

Social platforms are designed for engagement, not for our well-being.

I wanted to build a tool that does three things for me:
1. Raise awareness while scrolling - Giving me gently reminders when I'm slipping into doomscrolling patterns.
2. Analyze the content itself - run sentiment and thematic analysis on facebook, so I get insights into what kind of content is influencing my mood.
3. Block harmful/negative content - During the 2nd process, it's analyzing posts I didn't even see yet on my scroll view, so the harmful content was blurred out and it had a message about the type of content and an option to view it anyways.




Chrome extension base - hooks into the DOM while browsing.
Client-side analysis - I've been exploring lightweight ML models running in-browser (privacy-preserving, no server round trips).
Hybrid methods - in some cases, offloading analysis to a backend for deeper NLP tasks, but keeping the sensitive data on the client whenever possible.
Sentiment & topic analysis - the extension flags highly negative posts, patterns of repetitive themes, and gives the user a snapshot of what they've consumed in the last X minutes.



Frontend (Extension):

• Manifest V3 (Chrome's latest extension format)
  
• Content scripts injected into Facebook pages
  
• Popup UI built with vanilla JS and CSS (keeping it lightweight)
  
• Chrome storage API for persisting user preferences and analytics data
  

Analysis engine:

• TensorFlow.js for in-browser ML inference
  
• Sentiment analysis using a stripped-down BERT model (quantized to ~20MB)
  
• Layer of keyword extraction using TF-IDF implementation
  
• Real-time DOM mutation observers to catch new posts as they load
  

Backend - whitch is completely optional if you want to make it lighter:

• Node.js + Express for advanced analysis
  
• PostgreSQL for storing anonymized aggregate data
  
• Redis for caching frequent analysis results
  
• OpenAI API integration for deeper content analysis - still optional - you can only use Tensorflow if u want
  

When you first install the extension, it walks you through a quick setup:

• Set your "scroll time threshold" (default: 15 minutes)
  
• Choose notification style (subtle, moderate, or assertive)
  
• Select which topics to track (politics, news, entertainment, etc.)
  
• Privacy mode: client-only vs. hybrid analysis
  

As you browse Facebook, the extension silently works in the background:

What it does:

• Tracks scroll depth and time spent on page
  
• Analyzes each post you see (text content, not images yet)
  
• Builds a real-time sentiment profile
  
• Detects repetitive negative patterns
  

What you see:
A small toolbar that changes the highlighted color based on overall sentiment

• Green = mostly positive content
  
• Yellow = mixed or neutral
  
• Red = predominantly negative content
  

Hidden posts / images if the content is too negative - based on your onboarding selection.





When you hit your threshold (say, 15 minutes of scrolling with 70% negative content), you get a notification:
You can dismiss it, snooze for 10 minutes, or view your detailed dashboard.

Challenge 1: Facebook's DOM structure
Spoiler (Click to View)
Facebook uses heavily obfuscated class names and dynamic rendering. Posts don't have consistent selectors, so I had to:

Build a heuristic system that identifies posts based on DOM structure patterns

• Use multiple fallback selectors
  
• Implement a "learning mode" that adapts to FB's changes
  
• Monitor mutation observers for dynamically loaded content
  

Challenge 2: Sentiment analysis accuracy
Spoiler (Click to View)
Training data is biased, and social media language is weird:

Problems:

• Sarcasm breaks most models ("Oh great, another election post" is negative despite "great")
  
• Emoji analysis (❤️ vs 😡)
  
• Mixed languages in posts
  
• Context matters (a negative post about a good outcome: "Can't believe this corruption finally ended")
  

My solutions:

• Using emoji sentiment dictionaries as a supplementary signal
  
• Implementing context windows (analyzing 2-3 sentences together)
  
• Letting users flag incorrect analyses to improve the model
  
• Keeping confidence scores and only showing high-confidence results
  
• Using LLMs to make more accurate predictions
  

Challenge 3: Performance
Spoiler (Click to View)
Running ML models in-browser is slow:

Initial problems:

• Extension slowed down Facebook noticeably
  
• Battery drain on laptop
  
• Memory leaks from keeping too much data
  

Optimizations I made:

• Lazy loading: only analyze posts in viewport + 2 screens ahead
  
• Web Workers: offload analysis to background threads
  
• Model quantization: reduced model size from 85MB to 22MB
  
• Debouncing: batch analyze multiple posts together
  
• Caching: don't re-analyze posts you've already scrolled past
  

What I've learned

Technical lessons:

• Web scraping is fragile - always have fallbacks
  
• In-browser ML is viable but requires heavy optimization
  
• Performance debugging is 50% of the work
  

Personal lessons:

• Testing my own extension changed my scrolling habits
  
• Side projects don't have to be perfect to be valuable
  

Resources & References

If you want to build something similar, here are resources that helped me:

• Chrome Extension docs (Manifest V3)
  
• TensorFlow.js documentation
  
• Hugging Face model hub (for pre-trained models)
  
• DistilBERT for lightweight sentiment analysis
  
• OpenAI API for further analysis / better predictions
  

I'd love to get your feedback and if it's something you want to work on it, I'll prep the code and open source it



PS: "Quantum" styled text images were made by the awesome tool CL1B created

https://hackforums.net/showthread.php?tid=6310489