By Gaurab Chhetri - Research Project

CognitiveSky - Scalable Sentiment and Narrative Analysis for Decentralized Social Media

CognitiveSky is an open-source research infrastructure for analyzing mental health narratives on Bluesky, combining real-time ingestion, NLP pipelines, and an interactive Next.js dashboard. Accepted for presentation at HICSS 2026. Preprint available on arXiv, final version will appear in official proceedings.

CognitiveSky is a research platform and interactive dashboard for exploring mental health narratives on the Bluesky decentralized social network.
It was built to answer an important question: how can we track and understand large-scale discussions on mental health in real time?

Inspired by TwiXplorer ¹, CognitiveSky integrates streaming data ingestion, state-of-the-art NLP labeling, and interactive visualization into a single open-source framework. It has been accepted for presentation at HICSS 2026 (HICSS-59) ², highlighting its contribution to computational social science and mental health research.

Highlights

• Real-time ingestion of public mental-health-related posts via the Bluesky Firehose API.
• Automated NLP pipelines for sentiment, emotion, and topic detection using pre-trained transformer models.
• Daily reproducible snapshots stored in Turso (libSQL) and exported as JSON for transparency and reuse.
• Interactive Next.js dashboard for exploring narratives, topics, hashtags, and engagement trends.
• Fully open-source with reproducible pipelines, making it accessible to researchers and developers.

🖥️ Live Dashboard: https://cognitivesky.gaurabchhetri.com.np/

📑 Paper: arXiv Preprint | Accepted at HICSS 2026 (final version will appear in official proceedings)

System Overview

CognitiveSky is powered by two main components:

1. mh_worker (Data Ingestion)

• Node.js worker running 24/7 on Oracle Cloud free-tier VM
• Streams Bluesky posts using the Firehose API
• Filters for mental health keywords and stores posts in Supabase

2. summary.py (Labeling & Summarization)

• Python-based NLP pipeline triggered daily via GitHub Actions
• Performs:
  • Sentiment analysis (cardiffnlp/twitter-roberta-base-sentiment)
  • Emotion detection (j-hartmann/emotion-english-distilroberta-base)
  • Topic modeling (TF-IDF + NMF)
• Migrates labeled data to Turso DB
• Generates JSON summaries (sentiments.json, emotions.json, topics.json, hashtags.json, etc.)

These summaries power the interactive dashboard and allow temporal analysis of narratives.

Dashboard

The dashboard is built with Next.js + Tailwind + Recharts and provides:

• Topic timelines: How sentiment/emotion shifts across major discussions
• Hashtag & emoji analysis: Cultural and emotional signals in conversations
• Engagement insights: Most active users, posts, and community trends
• Narrative evolution: Tracking shifts in discourse over days and weeks

Impact

CognitiveSky demonstrates how AI, decentralized platforms, and open data can come together to support research in public health and online discourse. Its contributions include:

• A scalable, reproducible framework for analyzing social narratives
• A working example of 24/7 ingestion + NLP labeling + visualization
• A tool that researchers and advocates can extend to other domains beyond mental health

By combining technical rigor with accessibility, CognitiveSky bridges the gap between raw social media data and actionable insights.

Acknowledgements

• AIT Lab (Texas State University) ³: Mentorship and research guidance
• Bluesky Community: Open Firehose access and documentation
• Supabase + Turso: Database support for real-time and labeled data
• Hugging Face Transformers: Pre-trained models for sentiment and emotion detection

Citation

If you use CognitiveSky in your research, please cite our paper:

@misc{chhetri2025cognitiveskyscalablesentimentnarrative,
      title={CognitiveSky: Scalable Sentiment and Narrative Analysis for Decentralized Social Media}, 
      author={Gaurab Chhetri and Anandi Dutta and Subasish Das},
      year={2025},
      eprint={2509.11444},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.11444}, 
}

References