# Bartholomew Voice AI (BTP v5.4) — Outbound Cold Calling Engine An in-house, low-latency, speech-to-speech outbound cold calling system engineered specifically to sell **Bartholomew (BTP v5.4)** to engineering leaders, AI engineers, and tech executives deploying autonomous agents. --- ## 1. System Architecture ``` +---------------------------------------------+ | Bartholomew Voice Server | | (FastAPI + WebSockets) | +---------------------------------------------+ | | +---------------------------------------+ +---------------------------------------+ | | +--------------------------+ +---------------------------+ | Twilio Voice | | Browser Audio Bench | | (Telephony Carrier) | | (Web Audio Mic & Speaker)| | 8kHz ITU-T G.711 μ-law | | 24kHz PCM16 | +--------------------------+ +---------------------------+ | | +----------------------------> [ AudioCodec ] <-------------------------------------------------+ (mu-law <-> PCM16 24k) | v [ RealtimeVoiceSession ] (OpenAI Realtime API / Simulator) | +-----------------------+-----------------------+ | | v v [ Sales Persona: Alex ] [ Lead Queue Manager ] - Engineer-to-Engineer Pitch - Persistent JSON/CSV Storage - Sub-35μs AST Firewall Value - Automated Call Outcome Tracking - Objection Matrix (OpenAI, Cost, Integration) - Full Conversation Transcripts ``` --- ## 2. Quick Start: Test in 10 Seconds (Zero-Cost Simulation) You can test the conversational AI and objection responses right from your computer without spending a dime on API credits or telephony: 1. **Start the Voice Server**: ```powershell python run_voice_agent.py ``` 2. **Open the Web Audio Bench**: Navigate to [http://localhost:8765/voice](http://localhost:8765/voice) or the live hosted bench at [https://bartholomew.info/voice/](https://acn-26670.web.app/voice/). 3. **Click "Start Interactive Call"**: - Alex introduces himself with the engineer-to-engineer hook. - Click any objection button (e.g., *"Already have OpenAI guardrails"*, *"How much does it cost?"*, *"How hard to install?"*). - Listen to Alex's technical rebuttal and watch the live transcript stream in real time. --- ## 3. Connecting Real Outbound Calls (Twilio Carrier) To place real phone calls to prospects' mobile phones: ### Step 1: Get a Twilio Number ($1.15/month) 1. Sign up or log into [twilio.com](https://www.twilio.com). 2. Go to **Phone Numbers** → **Buy a Number** (costs ~$1.15/mo). 3. Copy your: - `TWILIO_ACCOUNT_SID` - `TWILIO_AUTH_TOKEN` - `TWILIO_PHONE_NUMBER` (in E.164 format, e.g. `+14155551234`) ### Step 2: Configure Environment Variables Add the following to your `.env` file or environment: ```env TWILIO_ACCOUNT_SID=ACXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX TWILIO_AUTH_TOKEN=your_auth_token_here TWILIO_PHONE_NUMBER=+14155551234 OPENAI_API_KEY=your_openai_api_key_here VOICE_PUBLIC_BASE_URL=https://your-domain-or-ngrok.ngrok-free.app ``` ### Step 3: Trigger Outbound Dialing To dial a lead via the REST API: ```bash curl -X POST "http://localhost:8765/api/dial?phone=+14155559876" ``` Or click **"Dial Real Phone Number"** directly in the Web Audio Bench! --- ## 4. Sales Script & Cold Calling Strategy ### Persona: "Alex" (Senior Infrastructure Engineer at Bartholomew) - **Opening Hook**: > *"Hey [Name], this is Alex from Bartholomew. Saw your team is building with autonomous agents at [Company] — did I catch you in the middle of something?"* - **Problem Formulation**: > *"We built Bartholomew — an open-source in-process AST firewall for AI agents. When agents have terminal or SQL tools, prompt injections or hallucinations can cause them to run destructive commands like `rm -rf` or drop tables. Have you guys run into tool-safety or spend runaway issues in your pipelines?"* - **The Value Proposition**: > *"It's a deterministic Python AST gate that blocks dangerous commands in under 35 microseconds before anything hits the OS. Installs with `pip install btp-guard` and wraps CrewAI or LangGraph with 1 line of code."* - **The Call-to-Action (CTA)**: > *"I'd love to send you our 2-page developer quickstart and the GitHub repo so you can inspect the code. What's the best email for you?"* ### Objection Handling Matrix | Objection | Alex's Proven Response | | :--- | :--- | | **"We already use OpenAI system prompts / guardrails."** | *"Totally hear you. System prompts are great for tone, but prompt injections easily bypass them. Bartholomew is deterministic syntax parsing right before `subprocess` or database execution — so even if the model hallucinates or gets jailbroken, the OS never receives the bad command."* | | **"How much does it cost?"** | *"The core engine is 100% free and open-source under Apache 2.0 via `pip install btp-guard`. We only have a $49/mo team plan if you need multi-tenant workspace keys or certified SOC 2 audit packages."* | | **"How hard is it to integrate?"** | *"Literally one line. You import `BTPTaskGuard` and wrap your existing agent tasks. Zero architecture rewrites."* | | **"I am too busy right now."** | *"Completely understand, I know you're in the middle of your day. Can I shoot you a quick 30-second link to `bartholomew.info` to check out whenever you have a minute?"* | | **"Send me an email."** | *"You got it. What's the best email for you? I'll send over the GitHub repo and the 1-line integration snippet."* | --- ## 5. Queue Management & Lead Tracking All outbound dialing state is persistently tracked in [`leads_queue.json`](file:///c:/Users/User/.gemini/antigravity/scratch/autonomous-circularity-network/leads_queue.json). ### Status States: - `PENDING`: Ready to be dialed. - `CALLING`: Dial in progress. - `CONNECTED`: Prospect answered. - `QUALIFIED`: Agreed to review docs or book engineering screen. - `NOT_INTERESTED`: Gracefully declined. - `VOICEMAIL`: Reached answering machine. - `DO_NOT_CALL`: Opted out. ### Adding Leads via CSV: Create a CSV with columns: `name,company,phone,email,role` and place it in the workspace, or use the `LeadManager.import_from_csv()` method.