CAIDEN

The problem

Small businesses miss 40–60% of inbound calls. The calls they do answer usually arrive during client meetings, deep work, or after hours. Hiring a full-time receptionist costs $35–50K per year and still doesn't cover evenings or weekends — leaving after-hours revenue on the table for any business whose customers make decisions outside 9 to 5. Existing IVR systems frustrate callers into hang-ups. Offshore answering services introduce accent and context friction, especially in healthcare or legal contexts where terminology matters. The market needed a voice AI that sounds natural, understands the caller's actual intent, holds context across a multi-turn conversation, and reliably books, qualifies, or routes — without sounding like a robot or dropping the call when the caller uses a phrase the script didn't anticipate.

Key challenges

Voice is unforgiving. A 2-second silence feels like a broken system; a wrong interpretation on turn 3 kills the whole call. Latency budgets are brutal — every hop between speech-to-text, model inference, and text-to-speech has to complete inside a human's turn-taking window (roughly 500ms end-to-end to feel natural), and speech synthesis has to be prosodically correct or callers immediately clock that it's a machine. On top of that, small business phone traffic is noisy — background sounds, thick accents, cross-talk, dropped words, callers naming things the system has never heard. Production voice AI has to gracefully recover from all of it without either freezing or hallucinating a plausible-sounding wrong answer.

What we built

CAIDEN is a production voice AI platform built for small-business phone traffic. The architecture separates the real-time audio path (streaming STT, low-latency TTS, edge-deployed inference) from the conversation orchestration layer (state machine, tool calls, business knowledge base). This split lets us optimize latency independently from reasoning quality. Inbound calls hit LiveKit for real-time audio streaming. Deepgram handles speech-to-text with a streaming interim-transcripts model that lets the orchestration layer start reasoning before the caller finishes speaking. ElevenLabs generates responses in a voice cloned per client so it matches their brand. A custom orchestration layer manages the conversation state machine, calls tools (calendar, CRM, knowledge base), and decides when to hand off to a human. The whole system runs on Fly.io with edge presence so audio never crosses a continent unnecessarily.

Our approach

1. 1

   Vertical prototype first

   Shipped a single-vertical prototype (a dental clinic) before generalizing. Real callers surfaced failure modes no synthetic testing would have caught — abbreviated names, phone-number readback errors, insurance terminology, patients handing the phone to family members mid-call.

2. 2

   Latency budget as a first-class constraint

   Set a 500ms end-to-end turn-taking budget and treated any change that broke it as a regression. This drove architectural choices — streaming STT with interim transcripts, prewarmed model contexts, edge deployment on Fly.io — rather than being optimization-phase afterthoughts.

3. 3

   Deterministic state machine over pure LLM

   The conversation flow is a state machine; the LLM handles natural-language understanding within each state. This prevents the model from hallucinating its way out of a booking flow — a common failure mode when agents are handed too much autonomy on structured tasks.

4. 4

   Human handoff as a first-class outcome

   Every call has a clean path to a human. The system knows what it doesn't know, transfers with full context (transcript, intent, caller data), and never pretends to be a person. This is what earned trust with the pilot clients.

Key architectural decisions

Results

• Handles 1,000+ calls/month for active clients across real estate, healthcare, and service businesses
• Sub-500ms turn-taking latency measured end-to-end (STT + inference + TTS)
• Zero critical incidents since production launch in 2024
• Renewals across pilot clients — retention is the true success metric for voice AI
• Multi-vertical platform: same core, per-vertical conversation flows (real estate, dental, medical, general service)
• Human handoff success rate — callers transferred with context rather than dropped or re-prompted

Impact

CAIDEN is the platform underneath our real-estate (Booked), healthcare (Clinic), and SMS deployments — every vertical we ship reuses the same orchestration, telemetry, and handoff primitives. That reuse is why we can vertical-launch a new deployment in weeks instead of months, and why the platform gets more capable with every client instead of accumulating divergent forks. The commercial outcome for pilot clients has been consistent: recovered after-hours revenue and reclaimed staff hours previously spent on phone triage.

Tech stack