FactLens

Inspiration

We were inspired by how quickly misinformation spreads across both written articles and video platforms, while verification tools remain fragmented and slow. Readers and viewers rarely take the extra steps to fact-check what they consume because it requires time, context-switching, and effort. We wanted to build a system that brings verification directly into the experience, whether someone is reading a news article or watching a video. FactLens was created around the idea that credibility analysis should happen instantly, in the background, and without friction.

What it does

FactLens is a browser-based verification tool that analyzes both written articles and video content. For articles, it extracts the main content, identifies 3-5 verifiable factual claims, searches trusted sources for supporting evidence, and evaluates each claim as true, false, or unverified with a confidence score. For video content, FactLens integrates TwelveLabs to analyze and retrieve relevant segments, enabling verification of claims made in multimedia formats. The system also provides an overall reliability rating and links to supporting evidence. To improve speed and efficiency, previously analyzed claims are cached using Valkey, reducing redundant processing and improving scalability.

How we built it

We built FactLens using a Chrome extension frontend and a FastAPI backend. The extension captures the full HTML of articles and sends it to the backend, where Mozilla Readability and BeautifulSoup extract clean, structured text across different website formats. We use Groq’s large language model to extract factual claims and compare them strictly against evidence retrieved from trusted sources via search APIs. For video verification, we integrated TwelveLabs to analyze and retrieve meaningful video content that can be evaluated alongside textual evidence. Claim verification is performed through carefully structured prompts that constrain the model to use only provided evidence. We implemented Valkey for caching verified claims using hashed identifiers, significantly improving performance and reducing repeated computation.

Challenges we ran into

One major challenge was handling the variability of article structures across websites. Simple DOM scraping proved unreliable, which led us to move content extraction entirely to the backend. Another challenge was ensuring consistent and structured outputs from the language model, which required defensive parsing and strict prompt engineering. Integrating video verification introduced additional complexity, as multimedia analysis requires different processing workflows than text. We also had to manage evidence relevance and prevent hallucinations by forcing the model to rely strictly on retrieved snippets. Coordinating the extension, backend API, LLM calls, search APIs, TwelveLabs integration, and Valkey caching required careful debugging and system design.

Accomplishments that we're proud of

We’re proud of building a full end-to-end verification pipeline that works across both text and video content. We successfully integrated backend article parsing, claim extraction, search-based evidence retrieval, video analysis through TwelveLabs, structured LLM verification, and scalable caching with Valkey into one cohesive system. We significantly reduced hallucination risks through constrained prompting and credibility scoring. Most importantly, we created a tool that empowers users to evaluate information in real time across multiple media formats.

What we learned

We learned that scalable content verification requires backend parsing rather than fragile frontend scraping. We learned that LLMs are powerful but must be tightly constrained to ensure reliable outputs. We also saw how multimedia verification introduces new architectural challenges compared to text-only systems. Search quality and source credibility directly impact the accuracy of verification results. Building a distributed system that integrates browser extensions, APIs, LLMs, caching layers, and video analysis taught us the importance of modular design, logging, and performance optimization.

What's next for FactLens

Next, we plan to improve source credibility modeling, refine multimedia claim detection, and enhance the accuracy of video-based verification workflows. We also aim to highlight verified claims directly within articles and video timelines for a more interactive experience. In the long term, we envision FactLens becoming a scalable credibility layer for the web, capable of analyzing news, video, and eventually social media in real time. Our ultimate goal is to make critical thinking seamless, accessible, and integrated into everyday information consumption.

Built With

• css
• groq
• javascript
• json
• python
• twelvelabs
• valkey