The Challenge
In an era of digital coercion and phishing, identifying threats in real-time is difficult. Traditional safety apps are reactive—they only help *after* you press a panic button. We needed a solution that was proactive, identifying threats via voice patterns and context without compromising the user's absolute privacy.
The Solution: Vibe Coding & Agentic AI
We built SENTIA, an AI agent that listens to call audio patterns to detect coercion. By leveraging Cerebras AI, we achieved sub-millisecond inference speeds, allowing the AI to "intuit" a threat instantly.
⚡ Speed (Cerebras)
Utilizing the Wafer-Scale Engine-3 to process audio tokens at record-breaking speeds (1,000+ TPS), ensuring no latency during live calls.
🔒 Zero-Knowledge Privacy
A core architecture choice: audio is processed ephemerally. Once the inference is made (Safe/Threat), the data is immediately discarded. No recording is ever saved.
Technical Architecture
- Orchestrator:Cline (GLM 4.7) handles the agentic workflow and logic.
- Inference:Cerebras Llama 3.1 70b model for pattern recognition.
- Voice:Cartesia for generating natural-sounding AI alerts.
- Frontend:Next.js, React, Tailwind CSS deployed on Vercel.
Project Demo
(Video auto-plays muted. Click controls to unmute)