Activity
Mon
Wed
Fri
Sun
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
What is this?
Less
More
1 contribution to Brendan's AI Community
Looking for an experienced AI Voice Agent / Realtime Voice Systems Consultant
I’m building a production AI receptionist for an appliance repair company. We already have a custom stateful backend with deterministic business logic, customer/service-request state, scheduling logic, corrections, multiple appliances, fees, service-area rules, and CRM integration work. The main unresolved problem is the realtime voice layer. We have worked with LiveKit and Retell and are currently evaluating Vapi. I’m looking for someone with hands-on production experience with Vapi, Retell, LiveKit, Deepgram, Twilio/SIP, or similar realtime voice stacks. The important problem is not prompt engineering. It is the boundary between a voice platform and an authoritative stateful backend: - finalized caller turns - transcript revisions - request retries / idempotency - interruptions / barge-in - assistant speech delivery authority - latency - streaming Custom LLM integration Initially I want a paid architecture/code review, not a rewrite. I’d like the consultant to review the current system and recommend the smallest reliable path to production, including whether Vapi, Retell, LiveKit, or another stack is the best fit. Please send: - voice platforms you have used in production - approximate call volume - examples of realtime voice problems you personally solved - experience with custom/stateful LLM backends - your hourly rate
2 likes • 26d
@Muhammad Shahbaz What exactly do you mean by “final caller turn” in Vapi, and what exact field do you use as the per-turn idempotency key?
1 like • 25d
@Muhammad Shahbaz Thanks, that makes sense, and the tool-call idempotency part is clear. The remaining case I’m trying to understand is semantic continuation after a pause. For example: Caller: “Yes.” [pause — Vapi produces a final transcript and the model emits the booking tool] Backend commits the booking. Caller then continues: “…but only after 4 PM.” Since the first tool call has already succeeded, toolCallId correctly prevents duplicate execution, but it cannot undo the fact that the action happened before the caller finished the larger thought. How do you handle that case in your production system? Do you rely on explicit confirmation before sensitive tools, make bookings provisional until another step, allow the next tool call to modify the previous action, or accept this as a bounded edge case?
1-1 of 1
Alec M
4
74 points to level up
@alexey-myachkin-9335
Appliance repair, smart home automations, ai dispatchers

Active 12h ago
Joined Oct 20, 2026
Powered by