New here · Brendan's AI Community

New here

Hey everyone — I'm Arpan, voice AI infrastructure engineer based in India. 28 years shipping production systems.

I build the layer most people don't touch — self-hosted LiveKit (not Cloud), SIP trunking, TTS/STT on bare-metal GPUs, telephony routing, failover, the works. When a VAPI project hits cost or scale limits, that's usually where I come in.

What's running in production right now:

— Enterprise voice system doing 300K+ calls/day, self-hosted telephony with SIP trunk integration

— Taxi dispatch: 10K+ daily calls, 99.3% success rate, 75ms address lookups, multi-language

— Self-hosted TTS across 10 GPU instances, 1.13M+ clips processed, cut a client's annual cost from $283K to $24K

— Real-time video avatar with controllable facial expressions (eyebrows, blinks, pupil movement) for a healthcare deployment

I've been running self-hosted LiveKit in production for a while now so happy to share what I've learned about the tradeoffs vs VAPI/Retell, especially around cost, latency, and scaling past the demo stage.

Looking forward to the conversations here.

3 comments