AI Cold Calling
Built for
Sales teams who want to call more prospects - without hiring more people
An AI voice agent that calls your prospect list, speaks naturally, answers questions, follows the campaign script and auto-classifies every call - with no human on the line.
What we built
A phone robot that actually talks - not just plays a recording
A typical call with the AI agent: it dials the number, delivers the opening greeting, listens to the prospect, understands what was said, responds intelligently and continues the conversation until a clear outcome is reached. The voice sounds natural - not robotic. Every conversation is transcribed and saved automatically.
- Automated outbound calls - the system works through the list without human input
- Real-time transcription via Deepgram (ro-RO and en-US)
- Responses generated by OpenAI GPT - fast and accurate, following the campaign script
- Natural voice via ElevenLabs - customisable per campaign
- Auto-classification after each call: Interested / Not Interested / Callback / DNC / No Answer
- Inbound calls - prospects who call your dedicated number are handled by the agent
- Test mode: type as the prospect and see the agent's replies, or run AI vs AI
- CSV list import, DNC management, live dashboard with per-campaign stats
Campaign setup
Define the agent once - it speaks consistently across thousands of calls
Each campaign has its own "personality": the agent's name, ElevenLabs voice, opening greeting, complete instructions on how to speak, product information and the interval between calls. You can create separate campaigns for different products or markets. A properly configured agent follows the script perfectly on every single call.
Call history
Every conversation transcribed and classified automatically
The calls page lists every conversation with its outcome status: called, interested, callback, DNC, no answer. Click any call to read the complete word-for-word transcript. Filter by campaign, outcome or date. No conversation is ever lost.
Prospect lists
Upload a CSV - the system calls from the first number to the last
Upload CSV files with phone numbers directly from the dashboard. You can manage multiple lists and assign them to different campaigns. The system automatically removes duplicates and checks DNC before the first call goes out. Lists generated by B2B Scraper import directly.
Test mode
Talk to the agent before going live with real calls
Test mode lets you simulate a call without dialling anyone. You type messages as if you're the prospect and the agent replies in real time - test how it handles objections, how it classifies the outcome, whether it sticks to the script. Or activate AI vs AI mode: a second model plays the prospect and converses autonomously with the agent.
Reports
Know exactly how many prospects were interested - and how many weren't
Per-campaign reports show: total calls made, how many answered, how many were interested, conversion rate, how many requested a callback. The main dashboard gives you a real-time overview - calls today, active campaigns, latest results. Decisions from data, not gut feel.
Why it matters
What changes versus a traditional call centre
Scale without headcount
A human agent covers 50-80 calls per day. The AI system can handle hundreds simultaneously, with no added staffing costs, no sick days and no variation in quality between agents.
Script followed 100%
The agent follows the campaign instructions exactly on every call - no improvising, no forgetting a key benefit, no bad days. Your message reaches every prospect consistently.
Full transcripts
Every conversation is saved as text. You can review it at any time, analyse recurring objections, refine the script, and understand exactly why prospects aren't converting - insights no traditional call centre can provide at this scale.
Complete pipeline
B2B Scraper collects the numbers → InboxIQ sends warm-up emails → AI Cold Calling calls the warm prospects. The three tools form a complete outreach pipeline, from raw data to real conversation.
Tech stack
Four AI services, one voice pipeline
Backend
- Python 3.12
- FastAPI + Uvicorn
- WebSocket (bidirectional audio)
- PostgreSQL + psycopg2
- JWT (python-jose)
- httpx / requests
- PyPDF / BeautifulSoup4
AI & Voice
- OpenAI GPT-4o-mini / GPT-4o
- ElevenLabs (Text-to-Speech)
- Deepgram (Speech-to-Text)
- ro-RO + en-US
- ulaw 8000Hz (telephony)
- MP3 44100Hz (browser)
Telephony
- Telnyx (digital carrier)
- TeXML (call control)
- WebSocket audio stream
- Outbound + Inbound
- Webhook events
Frontend
- React 18 + TypeScript
- Vite (build)
- Tailwind CSS
- Bootstrap Icons
- REST API + WebSocket
Notable architecture
- Bidirectional real-time audio over WebSocket between Telnyx and server - minimal latency
- Model escalation: gpt-4o-mini for fast replies, gpt-4o for complex questions mid-call
- TeXML instructs Telnyx to open a WebSocket audio stream immediately on call connect
- Post-call auto-classification: GPT reads the full transcript and determines the outcome
- AI vs AI test mode: a secondary model plays the prospect to test script handling automatically
- Per-agent knowledge base: PDF, DOCX or URL uploaded and parsed with BeautifulSoup / PyPDF
Running in-house on our own outreach campaigns - the same engine can be configured for your product and script, in Romanian or English.