Reality Check

Inspiration

In an era where generative AI can synthesize hyper-realistic images and audio in seconds, the "reality gap" has become a genuine threat to digital trust. We were inspired by the need to move beyond simple "vibes" when judging content authenticity. We wanted to build a tool that doesn't just guess, but uses agentic reasoning and real-time web context to prove whether what you're seeing is physically and contextually possible.

What it does

Reality Check is a browser-based "truth engine." Through a Chrome extension, users can capture any suspicious social media post. The content is then analyzed by an agentic pipeline that checks for visual artifacts and deepfake audio signatures. Most importantly, it performs a contextual "reality check"—using AI to search the web and determine if the scene depicted is even possible, such as verifying if a specific event actually occurred or if the laws of physics are being bent.

How we built it

We built a full-stack agentic solution using:

Frontend: A Chrome extension for instant visual and audio capture.

Orchestration: Airia to manage the workflow between Claude 4.5 (for contextual web-search reasoning) and Modulate’s Velma (for audio deepfake detection).

Backend: Google Cloud Functions (2nd Gen) to handle high-concurrency requests and data routing.

Data Layer: Google Cloud Storage for secure asset hosting and BigQuery for permanent metadata logging and historical analysis.

Analytics: Lightdash connected to BigQuery to give users a "lifetime trust dashboard."

Challenges we ran into

The biggest hurdle was balancing security with speed. We initially struggled with Public Access Prevention on our storage buckets; we had to ensure our AI agents could access the images for analysis without exposing sensitive user data to the open internet. Additionally, creating a full agentic workflow within a 4-hour hackathon window required tight architectural planning to avoid latency bottlenecks.

Accomplishments that we're proud of

We are incredibly proud of our Contextual Reasoning engine. While many tools look for pixel-level errors, our system can identify that an image is fake simply because the context—like an ant beating an elephant—is fundamentally false. We also successfully deployed a production-grade data pipeline that handles everything from raw capture to visual analytics in seconds.

What we learned

We learned that in the age of AI, "truth" is multimodal. Detecting a deepfake requires looking at the pixels, the audio frequencies, and the real-world context simultaneously. We also deepened our expertise in managing cloud permissions, specifically how to leverage Service Accounts and IAM roles to allow cross-platform tools like Airia to securely interact with GCP resources.

What's next for Reality Check

The next step for Reality Check is a closed feedback loop where the agent becomes more efficient with every scan. We plan to implement a system where the agent learns which 20 data sources are most reliable for specific content types, allowing it to bypass redundant searches and provide verdicts even faster. We also want to expand our detection suite to include real-time video stream analysis.

Built With

• airia
• gcp
• javascript
• lightdash
• python