Inspiration
I’ve always been interested in how news is framed, especially in politically sensitive regions like the Middle East and North Africa, where the same event can be presented in radically different ways depending on the outlet, the country, or the geopolitical context. While reading articles from various regional media, I often found myself asking: What am I not being told? How much of this is reporting, and how much is narrative construction? Most existing “media bias” tools focus on Western outlets, rely on static classifications, or provide oversimplified labels. They rarely analyze the article itself, its language, framing, omissions, and epistemic structure. That gap is what inspired me to build BonaFide.
What it does
BonaFide is a AI-powered lens that deconstructs an article into its raw components: claims, framing, and—most importantly—what was left out.
How we built it
Phase 1: The Foundation It started with a simple Python script using Gemini and newspaper3k to scrape an article and get a basic objectivity score. But I quickly realized that a single prompt wasn't enough to capture the complexity of modern journalism. Phase 2: The Multi-Agent Leap I moved to a Multi-Agent Architecture. Instead of one monolithic prompt, I built a pipeline of specialized AI experts: 1/ The Extractor: Focuses purely on the text, rhetorical techniques, and quotes. 2/ The Context Provider: Uses Ephemeral RAG (via Tavily) to pull in real-time external context, finding the "gaps" in the original article. 3/ The Comparator: Rigorously weighs the original article against the external context to identify framing and omissions. 4/ The Narrator: Synthesizes everything into a polished, human-readable report. Phase 3: Analytical Sophistication As the architecture grew, so did the features: • Genre-Awareness: The AI now understands the difference between a neutral News Report and an interpretive Op-Ed, adjusting its scoring logic accordingly. • Editorial DNA: Moving beyond labels to map articles to "Editorial Archetypes" like Western-Liberal or State-Aligned Nationalist. • Reader Risk: I added a cautionary field that warns: "If you read this without context, here is the misunderstanding you might walk away with."
Challenges we ran into
• Avoiding Hallucinated Context One of the hardest problems was preventing the model from inventing missing perspectives. This led me to design a controlled RAG pipeline and strict epistemic boundaries between: - Article-derived facts - External context - Analytical synthesis • Designing Meaningful Scores I had to rethink numeric precision and create transparent score breakdowns to avoid misleading users.
What we learned
This project became a deep learning experience across multiple domains -How to structure multi-stage reasoning pipelines -How to prevent hallucination through epistemic separation -How to design ephemeral RAG pipelines to structure counterfactual context using independent sources -Designing bias-aware prompting strategies This project fundamentally changed how I think about AI reliability, interpretability, and epistemic safety
What's next for BonaFide
Next step is moving to the video format analysis, a lot of news organizations are present on social media and present some of the news in the format of short videos or reels that rely heavily on emotion, music, cuts, captions which makes the risk of bias in them higher than in an article.
Log in or sign up for Devpost to join the conversation.