Veritas
Checking whether an image, its caption and the evidence tell the same story.
The problem
A real image can be paired with a false caption, recycled from another event or altered to look synthetic. A useful fact-checking tool needs to compare the media with its claim, explain what it can verify and be honest when its models or sources are inconclusive.
How it works
The Next.js client sends an image and claim to a FastAPI backend. A combined analysis runs image description, CLIP-based caption consistency, synthetic-image classification, an adversarial robustness suite and claim fact-checking in parallel, isolating errors so one failed stage does not hide the rest. The results bring these signals together to flag possible out-of-context media.
Fact-checking searches the Google Fact Check API and approved Goa news and government sources, then uses Groq to write a report with evidence and limitations. Report citations are checked against retrieved URLs, and uncertain or unsupported claims are not presented as proven false.
The robustness endpoint tests image and text edits plus single-step FGSM perturbations. The synthetic-image classifier is explicitly presented as a weak signal after near-chance results on held-out samples; the fixed robustness suite is not proof of general robustness. Reverse-image search is available as a separate endpoint and is not part of the combined web workflow.
Highlights
- First place at Hacksploit 2026 after Team Straw Hats topped the CTF round
- Combined image, claim, model and trusted-source analysis with per-stage status
- Source-attributed reports with citations checked against retrieved evidence
- Uncertain results and measured model limitations shown instead of overstated confidence
Screens
1st place