My Voice AI Agent Negotiated 800+ Business Deals in 1 Day (FULL Tutorial)

Key Concepts

• Voice AI
• Automation
• Vappy (Voice API platform)
• Prompt Engineering
• Tool Calling
• Lindy (AI for data organization)
• Arbitrage
• Shake and Stir (FCC regulation)

Voice AI for Lowballing Luxury Watch Dealers

Tony from SA Checkout discusses his project, "Lowballer 9000," a voice AI named Alex designed to negotiate prices for Rolex Daytonas with secondhand luxury watch dealers in the US. The goal was to automate the process of gathering critical details about the watch (condition, documentation, packaging) and then lowballing the dealers to find arbitrage opportunities.

• Target Audience: Secondhand luxury watch dealers (not official Rolex dealers).
• Objective: To find Rolex Daytonas at prices significantly below market value.
• Success Metric: Identifying dealers willing to sell at a low price, creating arbitrage opportunities.

Vappy: The Voice API Platform

Tony used Vappy (Voice API) to build and deploy Alex. Vappy is described as a flexible platform that doesn't lock users into proprietary systems.

• Key Features:
  • Allows users to bring their own prompts.
  • Supports various providers for different services (e.g., voice models).
  • Facilitates easy deployment of voice agents.
• Flexibility: Vappy allows users to choose their preferred voice models (e.g., Gemini 2.0 Flash, DeepSeek) and voices (e.g., Cartisia, New York Man).

Prompt Engineering and Iterative Improvement

The prompt for Alex was custom-written and iteratively refined based on real-world observations and test calls.

• Methodology:
  1. Observe real-world lowballing conversations.
  2. Simulate those conversations with Alex.
  3. Identify and correct undesirable AI responses.
• Example: Initially, Alex would rush through all questions at once. The prompt was modified to ensure concise, one-to-two sentence responses.
• Key Principle: Keep responses concise to minimize errors.
• Language Level: Tony found that using a sixth-grade English level in the prompt resulted in longer, more engaging conversations.

Model Selection and Temperature

The choice of voice model is crucial and depends on availability and performance.

• Model Selection: Tony initially used DeepSeek, but its API became overloaded, causing issues.
• Temperature: Controls the creativity of the AI's responses.
  • Temperature close to zero: More deterministic, predictable responses.
  • Higher temperature: More creative, but potentially less appropriate responses.

Tool Calling for Data Extraction and Integration

Vappy's tool calling feature allows the AI to interact with external tools, such as databases.

• Example: Alex used a tool call to record call transcripts and identify calls where an offer was made.
• Process:
  1. After each call, the transcript is analyzed.
  2. The AI determines if an offer was made, the lowball amount, and whether the offer was accepted or countered.
  3. This data is stored in an AirTable spreadsheet.
• Benefits: Enables efficient sifting through hundreds of calls to identify actionable leads.

Lindy Integration for Data Organization

Lindy, an AI tool for data organization, can be integrated with Vappy to further streamline data processing.

• Functionality: Lindy can automatically categorize and organize data in spreadsheets based on column names and content.
• Example: If a spreadsheet contains restaurant data and a "vegan friendly" column is added, Lindy can automatically check which restaurants offer vegan options based on their menus.

Overcoming Technical Challenges: Shake and Stir

A significant hurdle was ensuring that calls from the AI actually reached the intended recipients.

• Problem: Many calls went straight to voicemail due to FCC's "Shake and Stir" regulation.
• Solution: Register the phone number and complete KYC (Know Your Customer) verification to ensure calls are honored.

Use Cases Beyond Watch Dealing: Arbitrage Opportunities

Tony and Greg brainstorm potential applications of voice AI beyond the initial watch-dealing project.

• Real Estate: Identify potential sellers by asking targeted questions and analyzing responses.
• Framework: Use voice AI to gather data that provides an unfair advantage, enabling arbitrage opportunities.
• General Principle: Any business where knowledge and data are critical for success is a potential candidate for voice AI automation.

Startup Empire Feedback Bot

Tony developed a voice AI feedback bot for Startup Empire.

• Purpose: To collect feedback from members in a more natural and convenient way than traditional forms.
• Process: Members call a designated number and provide verbal feedback.
• Integration: Lindy summarizes the feedback and posts it to a Slack channel for the Startup Empire team to review.

Conclusion

The video demonstrates the power of voice AI for automating tasks, gathering data, and creating arbitrage opportunities. Vappy provides a flexible platform for building and deploying voice agents, while tools like Lindy enhance data organization and analysis. The key to success lies in careful prompt engineering, iterative improvement, and overcoming technical challenges like FCC regulations. The potential applications of voice AI are vast, particularly in industries where data and knowledge provide a competitive edge.