Negotiation Dojo

Master high-stakes salary negotiation with Viper, a ruthless AI trainer that sees your body language and hears your confidence in real-time.

Comment

🥋 Negotiation Dojo: The Story

💡 Inspiration

We've all been there: The job offer comes in. It’s slightly lower than you hoped. You know you should ask for more, but your heart races, your palms sweat, and you just say, "Okay, thanks!"

Salary negotiation is terrifying. Statistics show that 58% of candidates never negotiate, potentially losing over $500,000 in lifetime earnings. Reading blog posts doesn't help because negotiation isn't a knowledge problem—it's a pressure problem.

I built Negotiation Dojo to solve this. I wanted to create a "flight simulator" for difficult conversations—where you can get rejected, mocked, and lowballed by a ruthless AI, and then get expert coaching from Gemini 3 on exactly how to improve.

🤖 What it does

Negotiation Dojo is a dual-phase training system designed to build real-world confidence.

  1. The Simulation (The Pressure Test): You enter a real-time voice call with Viper, an aggressive AI negotiator. This phase uses high-speed, low-latency models to simulate the stress of a real argument. Viper hears your hesitation, interrupts you, and attacks your weak points instantly.
  2. The Analysis (The Intelligence - Powered by Gemini 3): This is where the real learning happens. Once the battle ends, the entire session transcript—including tonal markers and "silence" data—is fed into Google Gemini 3 Flash. Gemini 3 acts as your Senior Negotiation Coach. It doesn't just summarize; it performs deep psychological profiling:
  3. Detecting Anchoring Bias: Did you let Viper set the price too early?
  4. Silence Analysis: Did you rush to fill the awkward pauses (a sign of weakness)?
  5. Emotional Resilience: Did your vocabulary shrink when Viper got aggressive?

⚙️ How we built it

This project is a technical showcase of "Vibe Coding"—orchestrating multiple AI models to create a cohesive experience.

  • The Coach (Gemini 3 Flash): We utilized Gemini 3's superior reasoning and context window to process the chaotic, unstructured data of a live argument. Its ability to understand nuance—differentiating between "polite silence" and "scared silence"—is the core of our feedback engine.
  • The Simulator (Real-Time Audio): To create the immersive "Viper" persona, we integrated the Multimodal Live API via WebSockets. This handles the raw audio streaming and voice synthesis, serving as the data generator for Gemini 3's analysis.
  • Frontend Engineering: Built with React 18 and TypeScript, featuring a custom AudioWorklet pipeline that converts browser audio into the specific 16kHz PCM format required by the AI, all running client-side for zero latency.

🚧 Challenges we ran into

The biggest hurdle was coordinating the hand-off between the "Hot" (Real-time) and "Cold" (Analytical) AI systems.

  • Context Preservation: We had to ensure that the subtle cues detected during the live call (like long silences or interruptions) were accurately captured and passed to Gemini 3. We solved this by building a "Session Stats" aggregator that tracks metadata alongside the text transcript.
  • Audio Engineering: Handling raw PCM audio data in the browser (converting Float32 to Int16) was tricky, especially dealing with sample rate conversion to match Gemini's strict requirements.

🏅 Accomplishments that make us proud

  • Gemini 3 Integration: We successfully leveraged Gemini 3 not just as a chatbot, but as an analytical engine that provides actionable, high-level coaching advice.
  • Real-Time Latency: The application feels instant. You can interrupt the AI, and it reacts immediately, creating a truly immersive "Dojo" environment.

🧠 What we learned

  • The Power of Specialization: We learned that Gemini 3 shines brightest when given complex, unstructured data to reason over. By offloading the "chat" to a faster model and saving the "thinking" for Gemini 3, we got the best of both worlds: speed and intelligence.
  • Prompt Engineering for Coaching: We spent hours tuning the Gemini 3 system instructions to ensure the feedback wasn't just generic ("be more confident") but specific and tactical ("You anchored too high initially, then conceded $10k without getting a concession in return").

🚀 What's next for Negotiation Dojo

  • Deeper Gemini 3 Profiling: using Gemini 3 to generate a long-term "Negotiator Profile" that tracks a user's improvement over weeks of practice.
  • VR Support: Porting the experience to WebXR for true face-to-face intimidation.
  • Multiplayer Mode: Allowing a human to play as Viper against another human, with Gemini 3 serving as the live referee and judge.

Built With

Share this project:

Updates