Phoneline.ai
An AI agent that answers small businesses' phone calls from their own website and documents, and hands the call to a person when it should.
- My role
- Lead engineerBuilt the product end to end
- Timeline
- January 2025 to now
- Stack
- Python, FastAPI, PostgreSQLOpenAI Realtime API, React
- Live product
- phoneline.ai
The problem
Small businesses miss calls all the time. The owner is with a customer, the plumber is under a sink, the front desk is busy. The caller gets voicemail or hangs up, and the inquiry goes to a competitor.
Hiring a receptionist is expensive, and generic phone bots frustrate callers because they don't know anything about the business.
The idea
Most of what callers ask is already on the business's website: opening hours, prices, services, policies. Phoneline turns that content into an AI agent that picks up every call, answers like someone who knows the business, and passes the caller to a person when the question needs one.
Setup takes minutes: a business signs up, connects its website and any other documents, and forwards its calls.
What the product does
- Answers every callHandles common questions like hours, pricing and policies without anyone picking up.
- Knows the businessLearns from the business's website and documents, with no training sessions or scripts to write.
- Hands off to peopleTransfers the caller to the team when they need a person.
- Logs every callA dashboard shows each call, so the business can see who called and what they needed.
How it works
- Build the knowledge baseWhen a business signs up, its website and any other sources it provides are turned into a searchable knowledge base.
- Talk in real timeThe caller speaks with the agent through the OpenAI Realtime API, voice in and voice out, so the conversation flows naturally.
- Look up the answerWhen the caller asks something, the agent searches the knowledge base for the relevant passages.
- Answer from the business's contentThe reply is based on what the business has published, not the model's general knowledge.
- Hand off when neededIf the caller needs a person, the call is transferred to the team, and every call is logged in the dashboard.
What I built
- AI pipelineA RAG system over each customer's website and other sources, prompt design and response evaluation.
- Voice layerReal-time phone conversations on the OpenAI Realtime API, connected to the RAG system, plus call-transfer logic.
- BackendPython FastAPI services, PostgreSQL, third-party integrations, deployment and infrastructure.
- DashboardReact app for sign-up, agent configuration and call logs.
Engineering decisions
The hard parts of a phone agent are speed, accuracy and knowing when to stop.
Fast answers during a live call
Staying on the business's content
Turning websites into a knowledge base
Knowing when to hand off
Results
Need an AI feature like this?
Send a short description of what you're building and I'll reply with a concrete first step, usually within a day.