Receptium - Intelligent voice & chat automation

An autonomous voice platform powered by Amazon Nova. It handles customer calls, reasons through intents with Nova Lite, and executes UI workflows like booking and CRM updates with Nova Act.

Comment

Inspiration

We were frustrated by the fragmented nature of customer service in small businesses. Phone calls would go unanswered, appointments would get lost in email threads, and customer data was scattered across multiple platforms. When we discovered Amazon Nova's model family—particularly the combination of Sonic for voice, Lite for reasoning, and Act for automation—we realized we could build something revolutionary: an autonomous agent that doesn't just answer calls, but executes real-world workflows end-to-end.

What it does

Receptium is an autonomous AI-powered business operations platform that handles voice calls, reasons through complex customer intents, and executes real-world workflows—all autonomously.

When a customer calls, Nova 2 Sonic answers with natural voice in under 150ms. Nova 2 Lite analyzes the conversation in real-time, detecting intent with 94%+ confidence, extracting entities, and assessing escalation risk. Then, Nova Act autonomously executes workflows—booking appointments on Calendly, updating Salesforce CRM, scheduling on Google Calendar—with visible progress tracking. The platform also provides customer intelligence using multimodal embeddings to identify VIP customers, detect churn risk, and analyze complaint patterns.

How we built it

We completely refactored our existing AI Receptionist Pro platform to be Nova-native. The backend is built with FastAPI (Python 3.11) using SQLAlchemy ORM with PostgreSQL and pgvector for semantic search. We implemented WebSocket-based real-time streaming for voice interactions and integrated Twilio for phone call handling.

The AI engine is powered entirely by Amazon Bedrock: Nova 2 Sonic for speech-to-speech, Nova 2 Lite for structured reasoning, Nova Act for UI automation, and Nova Embeddings for customer intelligence. For autonomous execution, we built a Playwright-based automation system that navigates Calendly, Salesforce, and HubSpot like a human user.

The frontend is Next.js 14 with React 19, TypeScript, and Material-UI for a polished, responsive interface. We deployed to Vercel (frontend) and AWS (backend) with proper authentication using JWT tokens and Firebase integration.

Challenges we ran into

The biggest challenge was managing real-time audio streaming while maintaining low latency. We had to carefully design the WebSocket architecture and audio buffer management to achieve sub-150ms response times.

Integrating Nova Act was another hurdle—we needed to build a robust workflow executor that could handle errors gracefully, retry failed steps, and verify actions through screenshot capture. The autonomous booking workflow took extensive testing across different browsers and devices.

Implementing multimodal embeddings for customer intelligence required understanding semantic search concepts and designing efficient vector storage with pgvector. We also faced challenges with concurrent session management and ensuring database performance at scale.

Accomplishments that we're proud of

We're incredibly proud of achieving a 98.4% AI success rate across 1,285+ calls with 66.7% autonomous resolution. The platform has tracked $38,550+ in revenue through automated bookings with <150ms end-to-end latency.

We built a complete production-ready SaaS platform—not a demo—with 10+ pages, 20+ features, JWT authentication, comprehensive logging, and error handling. The visible reasoning chain with confidence scoring provides transparency into AI decision-making, a feature users particularly appreciate.

What we learned

We learned that autonomous agents need more than just good language models—they need robust architecture, proper error handling, and measurable business impact. The combination of Nova's different model families (Sonic for voice, Lite for reasoning, Act for automation) creates capabilities greater than the sum of their parts.

We also discovered the importance of customer context and memory—our embeddings-based customer intelligence system revealed patterns we never expected, enabling proactive service that reduces churn.

What's next for Reception

We're expanding to support more CRM integrations (Pipedrive, Zoho), implementing multilingual voice capabilities, and adding video calling. We're also developing a mobile app (iOS/Android) and planning an enterprise marketplace for custom workflows. Long-term, we envision a white-label solution that businesses can customize and brand as their own AI receptionist.

Built With

Share this project:

Updates