Sentinelle
A mobile prototype for checking claims in short videos in English, Chinese, Malay, and Tamil without leaving the feed.
Hackomania · 7 March 2026
Analysis in the Feed
We built Sentinelle to put claim assessments beside the videos people are watching. Viewers can open a badge to read supporting details, including information found through web searches of sources such as Singapore authorities and local news.
Sample videos and an anonymous demo mode let viewers explore the feed without creating an account. The verified, unverified, misleading, and disinformation labels are generated prototype assessments, not established verdicts on the videos.
From Video Claims to Supporting Sources
Turning the Video Into Claims
The React Native and Expo app uploads a video to Firebase Storage, then uses Gemini 2.5 Flash through Firebase AI Logic to transcribe it. Claim extraction turns that transcript into structured JSON fields. The analysis pipeline treats transcription and claim extraction as separate steps, moving from what was said to the statements to assess.
Adding Context Through Search
The assessment stage uses web search grounding to compare claims with supporting material, including Singapore authorities and local news. It produces a credibility score and supporting details. Those details give the viewer something to examine beyond the category displayed on the video.
Returning the Assessment to the Feed
Results are saved with the post in Firestore, and listeners update the feed as the data changes. The badge summarizes the result without taking over the video; opening it reveals the detailed assessment in a modal. This connects the compact feed indicator to the explanation behind it.
Team
Built with the Code and Conquer team.