The End of Deepfakes? How We Built a 6-Judge AI Ensemble Using Google Gemini
Disclaimer: This piece of content was created for the purposes of entering the Google Gemini Live Agent Challenge 2026.
In an era where synthetic media, deepfakes, and AI hallucinations are becoming indistinguishable from reality, how do we establish what is true? A single detection algorithm is no longer enough; model architectures evolve too fast, and synthetic generators are constantly finding ways to bypass traditional forensic filters.
To solve this, my team at Ekoahamdutivnasti Technologies built VASTAV Agent (Validation & Authentication System for Truth And Verification) — a radical new approach to deepfake detection powered entirely by the Google Gemini 2.5 Flash Lite model and Google Cloud infrastructure.
Reimagining Detection: The Multi-Agent Ensemble
Instead of training a monolithic neural network to spot a “fake,” we asked a different question: What if we put the media on trial in front of a panel of specialized, independent AI experts?
Leveraging the incredible speed and affordability of Google Gemini 2.5 Flash Lite, we designed a parallel orchestrator in Node.js that instantiates six unique AI “Judges” simultaneously when a user uploads media:
- The Forensic Analyst: Prompted to obsessively check for physical impossibilities—inconsistent lighting, impossible reflections, and incorrect sub-surface scattering.
- The Artifacts & Patterns Expert: Trained specifically on GAN signatures, diffusion markers, and algorithmic noise generated by tools like Midjourney or Stable Diffusion.
- The Contextual Analyzer: Built to scrutinize the semantic logic of the scene. Are the proportions correct? Does the structural engineering of the background make sense?
- The Physics Engine: Dedicated entirely to evaluating gravity, fabric tension, fluid dynamics, and material reactions.
- The Chief Justice: A macro-level overseer that evaluates psychological intent, emotional coherence, and potential manipulation tactics.
- The SynthID Detector: A specialized agent looking for the digital watermarks and metadata footprints pioneered by Google.
The Power of Multimodality
What makes Gemini the perfect engine for this is its native multimodality. We don’t just pass the image to the judges; we pass an entire context payload.
For instance, our Node.js Express server utilizes an exif-reader pipeline to strip hidden metadata directly from the user’s file format. We parse the Camera Make, Modification Date, and Software tags, and inject that raw text data directly into the prompt sequence for Judge 2. If an image claims to be a raw photograph but lacks basic camera EXIF data or exposes an “Adobe Photoshop” software tag, the Gemini model instantly weighs that into its contextual decision matrix.
The Consensus Engine
Once the 6 Gemini agents complete their independent analyses, our backend Consensus Engine tabulates the results. Because Gemini Flash Lite processes so quickly, we can run all 6 massive prompts in parallel and return a mathematically weighted final verdict to the user in seconds.
If any individual judge fails—perhaps due to a corrupted file read—the Consensus Engine seamlessly gracefully degrades, isolating the error and building the final verdict using the surviving agents without crashing the application.
Why Google Cloud & Gemini?
Building this required an API that was violently fast, fiercely intelligent, and aggressively cost-effective. Firing off 6 massive, highly detailed prompts for a single image upload would bankrupt a startup using traditional LLMs.
Because we built VASTAV on top of Google Gemini 2.5 Flash Lite, we achieved a near-instantaneous response time from 6 independent judges at a cost of practically zero (fractions of a cent per full scan).
The result isn’t just an app; it’s a scalable, impenetrable wall against synthetic misinformation.
Check out the open-source code on GitHub and learn how to build your own multi-agent ensemble!