ConflictCoach AI: The 'Flight Simulator' for Support

Real-time "Stress Meter" visualizing customer anger levels using ElevenLabs emotional telemetry.

Inspiration

Customer support is one of the toughest jobs in the world. Companies spend billions on training, yet new agents are often trained using boring PDFs, multiple-choice quizzes, or unrealistic roleplay scripts.

The problem? Reading about an angry customer is very different from actually being yelled at by one.

We wanted to bridge this gap. Just like pilots use flight simulators to practice crash landings without risking lives, we built ConflictCoach AI—a "Flight Simulator" for soft skills. It allows agents to practice handling the most difficult conversations (billing errors, wrong deliveries, technical glitches) in a safe, controlled environment before they ever pick up a real phone.

What it does

ConflictCoach AI is a web-based training platform where users enter simulated voice calls with AI agents that have distinct personalities and emotional triggers.

Real-time Conflict Simulation: Users speak into their microphone to handle scenarios like "The Wrong Delivery" or "The Billing Error."
Emotional Intelligence: Powered by ElevenLabs, the AI doesn't just read text; it yells, sighs, interrupts, and expresses genuine frustration. It dynamically reacts to the user's tone.
Live Stress Meter: A visual indicator shows the customer's anger level in real-time, gamifying the de-escalation process.
AI Coaching: After the call, Google Gemini 1.5 Flash analyzes the transcript to provide a detailed report card, highlighting empathy gaps, patience, and suggesting better responses.
Observability: We use Datadog to monitor the health of these real-time streams to ensure zero lag during training.

How we built it

We built a modern, low-latency architecture focusing on immersion:

Frontend: Built with Next.js 14 (App Router) and Tailwind CSS for a premium, glassmorphism-inspired UI. We used Framer Motion for the dynamic "Visualizer Orb" and "Stress Meter" animations.
Voice Engine (The Core): We utilized the ElevenLabs React SDK via WebSockets to create a full-duplex conversational stream. We engineered specific "System Prompts" to override the AI's natural tendency to be polite, forcing it to act angry, skeptical, or impatient until specific winning conditions (apology, refund offer) were met.
The Brain: We integrated Google Gemini 1.5 Flash via Vertex AI. We chose Flash for its speed and massive context window, allowing it to digest the entire conversation transcript instantly and act as a senior coaching mentor.

Challenges we ran into

Breaking the "Nice AI" Bias: LLMs are trained to be helpful and polite. Making the AI genuinely "rude" or "unreasonable" required extensive prompt engineering and temperature tuning.
Handling Interruptions: In a heated argument, people talk over each other. Achieving low latency so the user could interrupt the AI (and vice versa) was a technical hurdle we solved using the ElevenLabs Turbo v2.5 model.
Visualizing Voice: Syncing the visual orb animation and the "Stress Meter" with the AI's state required complex state management logic in React.

Accomplishments that we're proud of

Successfully creating an AI that feels "human" enough to raise the user's heart rate.
Building a beautiful, responsive UI that looks like a SaaS product, not just a hackathon prototype.
Integrating 6 distinct scenarios with unique personalities, proving the scalability of the platform.

What we learned

We learned that Latency is Emotion. In voice AI, a 2-second delay kills the immersion. Using ElevenLabs' low-latency Turbo models changed the experience from "talking to a robot" to "having a conversation." We also learned the power of Gemini's multimodal capabilities for rapid text analysis.

What's next for ConflictCoach AI

VR Integration: Bringing this into a virtual reality office environment.
Multi-language Support: Training agents in Spanish, French, and Hindi using ElevenLabs' multilingual models.
Enterprise Integration: Connecting with CRM systems (Salesforce/Zendesk) to train on real past ticket data.