Rep

The

Inspiration

Sales is one of the highest-paid professions, yet most sales reps get minimal practice before jumping on real calls with prospects. Athletes practice thousands of hours before game day—why shouldn't salespeople? We were inspired by a simple realization: the best way to get better at sales is to practice, but practice partners are expensive, inconsistent, and hard to schedule. What if AI could be your perfect practice partner—available 24/7, infinitely patient, and able to simulate any buyer persona? The idea crystallized when we saw the convergence of four technologies: Claude's advanced reasoning for realistic role-play ChromaDB + Claude for RAG system to retrieve product information in real time Real-time voice AI Agent with sub-500ms latency LLM-powered performance analysis We realized we could build something that doesn't just record calls—it actively makes salespeople better. How We Built It Architecture Overview We built a full-stack platform with three distinct AI pipeline stages: Frontend Stack: Next.js 14 with React 18 for the web app Tailwind CSS + shadcn/ui for beautiful, responsive design Vapi Web SDK for crystal-clear voice conversations Zustand for lightweight state management Backend Stack: FastAPI (Python) for async API endpoints Supabase (PostgreSQL) for data persistence with RLS ChromaDB for RAG-based hint system Vapi for voice AI orchestration The Three-Stage Pipeline Stage 1: Scenario Generation (Claude) When users input their product details and target persona, we use Claude 4.5 Sonnet to generate hyper-realistic buyer scenarios: The AI creates: Full persona backstory (role, company, pain points) 3-5 realistic objections they'll raise Hidden information (budget, timeline, competitors) Personality traits and communication style and much more Stage 2: Real-Time Conversation (Vapi + RAG) We integrated Vapi to handle the entire voice pipeline: Speech-to-text (Deepgram Nova-2) LLM orchestration (Claude as the buyer) Text-to-speech (ElevenLabs) Sub-500ms latency for natural dialogue RAG-Powered Hints: We built a hint system using ChromaDB that retrieves relevant sales information on the product the user is selling during the call: Stage 3: Performance Analysis (Gemini) After the call, we run comprehensive analysis: Core Analysis: Ge evaluates 7 skill categories (Discovery, Objection Handling, Rapport Building, etc.) with specific examples and timestamps. The system scores performance on a 0-100 scale per category: $$\text{Overall Score} = \frac{1}{7}\sum_{i=1}^{7} S_i$$ where $S_i$ represents each skill category score. What We Learned Technical Breakthroughs Voice AI is Production-Ready: Vapi's infrastructure gave us sub-500ms latency without building complex audio pipelines. This was game-changing—natural conversation requires instant responses. Voice AI is Production-Ready: Vapi's infrastructure gave us sub-500ms latency without building complex audio pipelines. This was game-changing—natural conversation requires instant responses. LLMs as Actors: Claude is shockingly good at roleplay when given detailed personas. The key is in the prompt engineering—we spent days iterating on the system prompt to make buyers feel authentic. LLMs as Actors: Claude is shockingly good at roleplay when given detailed personas. The key is in the prompt engineering—we spent days iterating on the system prompt to make buyers feel authentic. RAG for Real-Time Coaching: Embedding sales best practices in ChromaDB and retrieving relevant hints during calls was more effective than we expected. The context-aware suggestions feel like having a sales coach in your ear. RAG for Real-Time Coaching: Embedding sales best practices in ChromaDB and retrieving relevant hints during calls was more effective than we expected. The context-aware suggestions feel like having a sales coach in your ear. WebSocket State Management: Coordinating real-time transcript updates, audio streaming, and UI state required careful architecture. We learned to separate concerns and use React's concurrent features. WebSocket State Management: Coordinating real-time transcript updates, audio streaming, and UI state required careful architecture. We learned to separate concerns and use React's concurrent features. Product Insights Feedback Must Be Specific: Generic feedback like "good job" doesn't help. We had to timestamp exact moments and quote conversations to make insights actionable. Feedback Must Be Specific: Generic feedback like "good job" doesn't help. We had to timestamp exact moments and quote conversations to make insights actionable. Difficulty Calibration Is Hard: Making scenarios progressively harder while keeping them realistic required extensive prompt engineering and testing. Difficulty Calibration Is Hard: Making scenarios progressively harder while keeping them realistic required extensive prompt engineering and testing. Users Want Progress Tracking: The analytics dashboard became as important as the practice itself. Seeing improvement over time drives engagement. Users Want Progress Tracking: The analytics dashboard became as important as the practice itself. Seeing improvement over time drives engagement. Challenges We Faced Challenge 1: Audio Latency Battle Problem: Initial experiments with custom WebRTC + OpenAI Realtime API had 1-2 second delays, killing conversation flow. Solution: Switched to Vapi, which handles the entire audio pipeline with optimized edge infrastructure. Latency dropped to <500ms, making conversations feel natural. Challenge 2: Transcript Synchronization Problem: Real-time transcripts from Vapi arrived out-of-order, causing UI jumps and confusion. Solution: Implemented event sequencing with timestamps and a reconciliation algorithm: Challenge 3: Cost Optimization Problem: Running three Claude API calls per session (scenario, conversation, analysis) was expensive (~$2-3 per session). Solution: Caching common scenarios (reduced scenario costs by 70%) Using Claude Haiku for simple tasks Batching analysis requests Implementing strict session duration limits Final cost: ~$1.00-1.85 per 15-min session Challenge 4: Feedback Problem: Initial analysis was too vague: "You did well on discovery." Solution: Structured output with: Timestamped examples from transcript Specific quote + feedback pairs Side-by-side comparison of what was said vs. what could've been said Quantified metrics (talk time ratio, questions asked, objections handled) What's Next We're just getting started. Future plans include: Multi-persona calls (practice with buying committees) Voice cloning (practice with recordings of real buyers) Team features (manager dashboards, shared scenario libraries) CRM integration (import real prospect data) Mobile app (practice anywhere) The Big Lesson Building AI products is less about the models and more about the orchestration. The magic isn't in one API call—it's in chaining multiple AI services together with great UX to create something that feels seamless. Rep proves that with the right architecture, AI can be more than a chatbot—it can be a coach, a practice partner, and a path to mastery.