Sure — here’s a clean, hackathon-submission ready writeup for each section (crisp but solid). You can paste this directly.
Inspiration
With the rapid growth of digital media, users are overwhelmed by hundreds of news headlines daily. Most platforms either show generic trending news or personalize feeds without explaining why content is shown. At the same time, LLM-based summarization often generates hallucinated information, making users hesitant to trust AI-generated news. This inspired us to build a system that is personalized, trustworthy, and explainable.
What it does
Agentic AI–Driven Personalized News Summarization is a platform that:
Learns user interests from onboarding + reading behavior
Retrieves the most relevant news content using embeddings and a vector database
Generates fact-grounded summaries using RAG to reduce hallucinations
Uses multiple agents for retrieval, personalization, summarization, and verification
Provides source citations and “Why this news?” explainability
Continuously improves through user feedback (“More/Less like this”)
How we built it
We built the system using a modular ML pipeline:
News ingestion via RSS feeds / APIs / scraping
Cleaning, deduplication, chunking, and metadata extraction
Sentence-transformer embeddings for article chunks
FAISS/Chroma vector database for semantic retrieval
User interest embedding created from user interactions
Agentic pipeline:
Retrieval Agent → fetches relevant chunks
Personalization Agent → aligns content with user intent
Summarization Agent → generates grounded summary
Bias/Verification Agent → cross-checks across sources
Frontend dashboard to display summaries, citations, and feedback controls
Challenges we ran into
Reducing hallucinations while keeping summaries short and readable
Designing personalization for cold-start users with no history
Balancing retrieval relevance with recency and diversity
Keeping multi-agent workflow fast enough for real-time use
Ensuring explainability is clear but not overwhelming
Accomplishments that we're proud of
Built a working RAG-based summarization pipeline grounded in retrieved sources
Implemented user-interest based personalization that changes what is retrieved
Designed a multi-agent architecture instead of a single monolithic model
Added explainability (“Why this?”) and feedback-based adaptation
Created a scalable design that can grow into a full production system
What we learned
Personalization becomes more powerful when applied at the retrieval stage, not just ranking
RAG significantly improves trust by grounding summaries in real content
Agentic decomposition improves modularity and makes debugging easier
Cold-start is a major challenge and needs explicit onboarding signals
Explainability is critical for responsible AI in news applications
What's next for Agentic AI–Driven Personalized News Summarization
Add stronger bias detection using multi-source comparison and stance analysis
Improve personalization using session-based recommendation models
Add multilingual support and region-based news personalization
Integrate credibility scoring and fake-news signals
Deploy as a scalable web app with real-time indexing and caching
Log in or sign up for Devpost to join the conversation.