Bivium Veri AI

Bivium Veri AI: A real-time multimodal profiler that decodes the split-second choice between truth and deception using Gemini Live API 🕵️‍♀️

Comment

💡 Inspiration

The inspiration came from the idea of creating a digital "lie detector" with a soul. We wanted to build an AI that doesn't just process text, but observes the human behind the screen—their hesitation, their facial expressions, and their tone. That's how Agent Elena was born: a sharp-tongued, sarcastic psychological profiler.

⚙️ How we built it

The project is built on a modular Python architecture:

  • Core Logic: A multi-threaded system that handles simultaneous video capture, audio processing, and AI inference.
  • Vision: OpenCV for real-time video stream processing and UI overlay (HUD).
  • Brain: Google Gemini 2.0 Flash (Multimodal), capable of "seeing" the user through frame analysis.
  • Voice: A dedicated TTS worker using pyttsx3 with a custom-tuned voice profile to match Elena's personality.
  • Resilience: A custom "Fallback Engine" that automatically cycles through different Gemini models (2.0 Flash, 2.0 Lite, 1.5 Flash) if API limits are hit.

🧠 What we learned

We learned how to manage complex threading in Python to prevent the UI from freezing during heavy AI requests. We also mastered multimodal prompting—teaching the AI to correlate visual cues (like "looking tired") with spoken words.

🚧 Challenges we faced

The biggest hurdle was the "429 Resource Exhausted" challenge. During the hackathon's peak, API quotas were extremely tight. We solved this by implementing a smart retry logic and a model priority queue. We also fought hard with Windows COM interfaces for stable voice output in a multi-threaded environment.

🚀 Accomplishments that we're proud of

We are proud that Elena actually feels "alive." She isn't just a chatbot; she's a critic. Seeing her successfully analyze a user's micro-expressions and respond with a sarcastic remark in real-time was our "Eureka" moment.

Built With

Share this project:

Updates

posted an update —

Update: Agent Elena is officially LIVE! What a journey! From a blank script to a fully functional multimodal AI profiler. Here is how Bivium Veri AI evolved during this intense session:

Multimodal Brain: Successfully integrated Gemini 2.0 Flash to analyze video frames and audio simultaneously.

The "Elena" Personality: Tuned the AI prompt to create a sarcastic, sharp-tongued psychological observer. She doesn't just talk; she judges!

Custom HUD: Built a custom Python overlay using OpenCV and Pillow to display real-time status, audio waveforms, and AI verdicts directly on the camera feed.

Stability First: Implemented a robust "Fallback Engine" that automatically cycles through multiple Gemini models to beat the "429 Resource Exhausted" errors.

Check out the screenshot of our final HUD below! This is the result of 20 years of technical experience meeting the cutting edge of AI.

Log in or sign up for Devpost to join the conversation.

Submission history