The problem
The vendor voice platform proved the concept, but at about 10¢/minute, scaling AI interviews to thousands of monthly sessions made unit economics painful. Voice was becoming the single largest variable cost in the product.
The build
Neeraj engineered a custom voice-agent stack on LiveKit: WebRTC transport, streaming STT, per-stage LLM routing, and low-latency TTS via SmallestAI and Cartesia. Committed-use infrastructure and provider credits pushed marginal cost toward zero; the steady-state figure landed at ~2.5¢/minute all-in.
The credits on top
Beyond engineering the stack down, we found a second lever most teams miss: infrastructure credit programs. We identified that the LiveKit workload qualified for a significant credit opportunity, pursued it, and secured up to $100,000 in LiveKit credits applied to eligible voice AI infrastructure. While those credits were available, the cost of that eligible infrastructure ran at $0. We are careful about how we frame this: the credits made those workloads free while they lasted, not permanently. Even as they draw down, the underlying stack still runs near 2.5¢ per minute, so the unit economics stay strong either way.
The result
75% per-minute cost reduction with interview quality metrics unchanged. This stack, and the honest 'when to use a platform vs when to go custom' decision framework, is exactly what Axionry's Voice AI Development service delivers to clients.