The problem
Businesses miss revenue at the phone: sales calls go unanswered or get an untrained pitch, and front desks can't cover every inbound call while also running the room. Staffing 24/7 coverage that stays on-message is expensive and inconsistent.
A voice agent that can actually do this has to sound natural with real turn-taking and barge-in, stay tightly on-script for pricing and product claims (without inventing numbers), handle objections, and complete the real action — book the appointment, send the confirmation, route the call.
The approach
We built a set of real-time voice agents across the two dominant delivery modes — browser (WebRTC) demos and full telephony — reusing the same pattern: a carefully engineered persona and knowledge base, tool calls for the concrete actions, and guardrails against hallucinated pricing or specs.
One agent is a browser-based sales rep for an industrial laser-marking manufacturer that recommends the right product line by use case, quotes pricing only within approved floor-and-up ranges, deploys financing and objection-handling plays, and hedges to a specialist rather than inventing a spec. Another is an inbound service-department advisor for a multi-location auto dealership that books, looks up, and modifies service appointments and sends SMS/email confirmations. A third is a small-business AI-receptionist platform where owners configure their own agent (business, services, hours, voice) and watch live call analytics.
Secrets stay server-side (the browser gets only short-lived, gated access tokens), and behavior is locked down with automated behavioral test harnesses rather than vibes.
What I built
- Browser voice sales agent (WebRTC) with a voice-adapted sales prompt + inlined knowledge base, server-minted short-lived access tokens, and passcode-gated call minting so provider keys never reach the client
- Automated sales-behavior test harness asserting product recommendations, floor-and-up pricing discipline, objection handling, spec-hedging, and barge-in
- Inbound dealership service-advisor agent with create / look-up / modify booking tools, SMS and email confirmations, and multi-tenant per-dealership config (hours, services, persona)
- Configurable AI-receptionist platform: live agent provisioning from business config, call-completion webhooks, LLM transcript analysis and outcome classification (phone collected, appointment scheduled), and a real-time WebSocket analytics dashboard with per-session isolation
- Natural turn-taking and barge-in across all agents, with guardrails that refuse to diagnose or quote outside approved bounds
- Retell
- ElevenLabs
- OpenAI GPT-4o
- VAPI
- Twilio
- Fastify
- Express
- React
- PostgreSQL
- TypeScript
- Railway
Result
A reusable playbook for sales and reception voice agents spanning browser and telephony, provider-agnostic across the major real-time voice stacks and grounded to a client's real products, pricing rules, and booking systems.
- Multiple agents deployed to production hosting, each grounded in a specific business's knowledge base and rules
- Behavioral correctness enforced by automated test suites, not manual spot-checks
- Concrete actions wired through: real booking create/modify/lookup, SMS/email confirmations, and outcome-classified call analytics