Elyx Ai

⁠Elyx: The AI that answers calls, runs operations, and never clocks out built for service businesses.

Comment

Elyx: The AI Operating System for Local Service Businesses

Inspiration

Every day, thousands of small businesses garages, dental practices, salons, restaurants lose revenue to a problem as old as the telephone: missed calls.

A single missed call at a busy dental practice can mean €200+ in lost revenue. For a garage, it's a job that goes to a competitor. For a restaurant, it's an empty table on a Friday night. Multiply this across millions of service SMBs, and you're looking at billions in lost economic activity.

I built Elyx because I saw this problem firsthand. Local businesses don't need another SaaS dashboard they need an AI that actually does the work: answers the phone, books the appointment, follows up with the customer, and never takes a day off.

What I Learned

Building a production voice AI system taught me that the last 10% is 90% of the work. Getting an LLM to have a conversation is easy. Getting it to:

•⁠ ⁠Sound natural with regional accents (Irish, British, German) •⁠ ⁠Handle interruptions and crosstalk gracefully •⁠ ⁠Integrate with legacy booking systems via webhooks •⁠ ⁠Recover from misheard names and phone numbers •⁠ ⁠Know when to escalate to a human

...is where the real engineering lives.

I also learned that SMBs don't buy "AI" they buy outcomes. "We use Gemini" means nothing to a garage owner. "You'll never miss another call" means everything.

How I Built It

Architecture: •⁠ ⁠Voice Layer: Gemini Live API for real-time, multi-turn conversation with sub-200ms latency •⁠ ⁠Orchestration: Pipecat for managing audio streams, VAD, and turn-taking •⁠ ⁠Business Logic: Custom workflow engine that maps conversations to structured entities (customers, bookings, jobs) •⁠ ⁠Integrations: Webhooks to existing POS, calendar, and CRM systems per vertical •⁠ ⁠Infrastructure: Deployed on GCP with EU data residency (GDPR compliance is non-negotiable for European SMBs)

Vertical Modules: Each vertical (garage, restaurant, salon, dental) has a purpose-built module with domain-specific: •⁠ ⁠Vocabulary and entity extraction •⁠ ⁠Workflow templates •⁠ ⁠Integration mappings

The AI doesn't just "chat" it operates within a structured state machine that ensures every conversation results in a concrete business outcome.

Challenges

  1. Latency vs. Intelligence Trade-off Users expect human-like response times (<300ms). But smarter responses require more context and reasoning. We solved this with streaming responses and predictive pre-computation of likely next turns.

  2. Accent and Dialect Handling Irish English alone has dozens of regional variations. Names like "Siobhán" or "Caoimhe" need phonetic handling. We built a custom pronunciation layer and fine-tuned for local speech patterns.

  3. Integration Hell SMB software is fragmented dozens of POS systems, booking tools, and CRMs, many with no API at all. We had to build scrapers, email parsers, and even IVR bridges for legacy systems.

  4. Trust Getting a business owner to let an AI answer their phone is a leap of faith. We built confidence through transparency: real-time call monitoring, instant escalation, and detailed transcripts.

What's Next

Elyx started as an AI receptionist. It's becoming the AI operating system for service businesses—owning not just the call, but the entire customer lifecycle: booking → service → payment → retention.

The vision: one AI layer that understands your customers, orchestrates your operations, and grows your business whether you're a one-chair barbershop or a 50-location dental group.

Built With

Share this project:

Updates