GoIPpro
Behind the Glass
Live deployment journal

Behind the Glass

A real internet provider in Ukraine. Real customers calling in. A voice AI named Asya answering the phone. This is the unedited journal of teaching a machine to talk — published as it happens, bugs included.

No case-study gloss. When Asya loses a customer, we write that down too. That's the point of the glass: you get to watch.

Dispatch 1 · July 4, 2026

Seventeen Slash Twelve

A woman wanted internet and was turned away — though coverage was there all along. The culprit: our own assistant's tidiness instinct, quietly deleting the slash from a house number. A detective story about one lost call, and the fix that gave every corner building its second chance.

Dispatch 2 · July 7, 2026

Eighteen Seconds into an Empty Room

The “truncated” call recordings turned out to be intact: our voice assistant kept talking into an empty room for eighteen seconds after the caller hung up — and we were paying for every one of those seconds.

Dispatch 3 · July 9, 2026

The Ghost of the First Second

Sometimes our voice assistant went silent for an entire call, if the caller said “hello” too fast. How we tracked the ghost through her inner draft — and taught her to cover her ears.

Dispatch 4 · July 11, 2026

Fifteen for One

Fifteen human support operators couldn’t work through the blackouts — one voice agent replaced them. An honest story about what already works, what is still rough, and how it scales.

Dispatch 5 · July 13, 2026

The Voice We Fired

The opening of our mini-series on Asya’s six voice engines. First up — ElevenLabs: the most beautiful voice on the market, fired not for its bugs but because its character was edited with a mouse in a browser.

Dispatch 6 · July 15, 2026

The Price of Perfection

Asya’s second engine — OpenAI Realtime. There was almost nothing to fault in the quality; there was plenty to fault in the bill: 22.6 cents per minute of conversation. A story about being defeated not by technology but by arithmetic.

Dispatch 7 · July 17, 2026

Five Watchdogs

The Gemini pilot lasted twelve days. To keep the engine from simply dying mid-call, we had to post five watchdog timers around it at once. A chronicle of our most capricious candidate.

Dispatch 8 · July 19, 2026

The Shortest Romance

We closed the Grok pilot in under two days: three scenarios out of ten, a stubborn slide into the wrong language, and tools it refused to call. Plus the quiet hero on the bench — both stories in one dispatch.

Dispatch 9 · July 21, 2026

The Sixth Voice

The finale of the mini-series: why Asya’s brain became Ultravox — an engine without the loudest name, but with a countable price, a visible transcript, and honest limitations we know by name.

Dispatch 10 · September 21, 2026

One for Three

The contact center grew from one support person to three, but Asya, our Iron Lady, is still on her own — with something new under the hood. Ultravox has retired and Asya runs on Soniox: why, what got better, and what it still cannot do.

Dispatch 11 · October 4, 2026

Six Hundred Forty-Six and One

646 languages is our target, not a status line. European models are in training now, African and Arabic are next, and the real bottleneck is GPU capacity — we and our partners have requests in for more. First up: continental Portuguese, full duplex, on our own infrastructure.

Dispatch 12 · October 5, 2026

It Is Not Only Customers Who Call at Night

Running a voice platform on your own servers means defending it. Scanners now hunt for AI agents with tools attached, and the nastiest attacks aim at the agent itself, not the server. What the 2026 research shows, and the rules we work by.

Dispatch 13 · October 6, 2026

Open voice AI: what has already happened in practice

An open voice AI button is an open paid resource. Seven documented cases — from a bot swearing at its own company to millions of leaked audio recordings — and the limits we build in before a public launch.