Callin.io · production voice AI
Callin.iovoice AI case study
How Callin.io scaled voice agents serving 500+ businesses, cut LLM cost roughly 20%, and designed telephony and billing around latency. 9M+ conversation minutes processed (per Callin.io).
This is the canonical write-up of the work: routing each turn by complexity, a latency budget on the call path, and post-call analytics that show what to fix. If you are hiring an AI Voice Engineer or technical lead, start here, then book 20 minutes.
- businesses served
- 500+
- lower LLM inference cost
- ~20%
- fewer cancellations
- ~40%
One live turn, with the intent, model and cost for that turn.
Callin.io is a voice AI platform. Callin.io is pre-incubated by I3P, the incubator of Politecnico di Torino. I started as a Software Engineer building it from the ground up and now lead its engineering as Technical Lead. This page covers the problem, how I approached it, and what changed.
(01)Problem
Role: Software Engineer, then Technical Lead, at Appspundit Infotech. Timeline: October 2024 to present. Callin.io is used by businesses in healthcare and real estate, among others. On a voice call, a slow answer is a hangup, and the cost of every turn adds up.
- Every turn of a voice call is a race. If the agent takes too long to answer, the caller hangs up.
- Using one large model for everything burned margin on simple turns like greetings and confirmations.
- Customers were leaving, and the causes were spread across latency, voice quality, language support and how hard agents were to manage.
(02)Approach
Callin is a multi-tenant voice-agent platform. Customers build agents, launch campaigns, attach phone numbers and manage everything from one dashboard. Healthcare and real estate show up as customers and templates, not as certifications.
The builder, the campaign tools, and the numbers are the product they click. Under that, every call still follows the same path: telephony in, orchestrator, models, voice out, and post-call analytics. I am not dropping a fake screenshot of the builder. You can see the product at callin.io. This page is the system under it.
Route by complexity
Classify each turn and send it to the smallest model that can handle it, keeping large models for hard reasoning. This is what cut inference cost by roughly 20%.
Treat latency as a budget
Each part of the call path gets a time allowance, and anything that blows it gets fixed first.
Choose providers on evidence
I compared speech and voice providers side by side and recommended a switch, which the CEO approved.
Learn from every call
Post-call analytics (goal analysis, sentiment, key moments, recommendations) show customers what to fix in their agents.
(03)Voice stack
The production path has five parts: the caller, telephony, an orchestrator, the models, and post-call analytics. That is the ownership map, not a runbook.
The orchestrator decides which model handles each turn and enforces the latency budget. Simple turns go to small models. Hard reasoning goes to large ones. Cost and speed are visible on every turn.
After the call, self-hosted models analyze goals, sentiment and key moments, so customers can see what to fix without sending call data to extra vendors.
- 01CallerPhone or web
- 02TelephonySIP · WebRTC
- 03OrchestratorRouting · cache · latency budget
- 04ModelsSTT · LLM · TTS
- 05Post-call analyticsInsights back to the customer
We own the path end to end. Cost and latency are visible on every turn.
(04)Outcome
Three decisions that mattered. Problem, decision, outcome.
A.Routing by complexity
I used one large model for every turn. Greetings and confirmations burned margin.
I classify each turn and route it to the smallest model that can finish it.
Inference cost is roughly 20% lower.
B.Choosing providers on evidence
Voice quality and latency varied between speech and voice providers. Opinions were not enough.
I compared providers side by side on the same calls and recommended the best fit.
The CEO approved the switch.
C.Letting AI agents fix tickets, with limits
Engineering time was going into small, repetitive support tickets.
I let AI agents handle fixes, gated by how much damage a bad change could do. Low-risk fixes go through. Risky ones wait for a human.
Engineers spend their time on harder problems. No change ships without matching its risk level.
Three numbers, because this page is supposed to earn them.
500+ businesses served, roughly 20% lower inference cost, and roughly 40% fewer cancellations.
Next I want tighter per-turn latency visibility, so a slow path cannot hide behind a good average.
If you are hiring a technical lead who has owned a voice platform's architecture and cost, this is the work.