# How Personal AI's Miroslav Volek Is Moving AI Into the Telecom Network > Source: > Published: 2026-11-14 01:52:08+00:00 # How Personal AI’s Miroslav Volek Is Moving AI Into the Telecom Network ## The Personal AI engineering chief is developing distributed voice-AI systems designed to reduce latency and computing costs across telecom networks. Earlier this month at NVIDIA GTC, telecom operators showed what becomes possible when AI runs inside their networks, close to the people who use it. One of the projects NVIDIA featured as a benchmark case study was a distributed inference architecture for voice AI. In a Comcast benchmark, it reached 42,362 tokens per second, an 80.9% gain, and cut the cost per token by 76.1%, while voice replies stayed under 500 milliseconds. In practical terms, the same GPUs produce nearly twice as much output, each answer costs about a quarter of what it did, and a caller hears the reply without a noticeable pause. The architecture was designed by Miroslav Volek, Head of Engineering at [Personal AI](https://www.personal.ai/). For Volek, the result is the latest step on a long road that began with a question he first asked as a boy: what does a machine actually do when a person uses it? He reached AI through machine tools and government software, teaching himself each new field along the way. ## From programs on paper to a medical AI grant Miroslav Volek first saw a computer at 11, at his cousin's house. His cousin moved a joystick, and the figure on the screen moved with it. The game was simple, but the exchange between a person and a machine stayed with him. He had no computer of his own, so he bought a book on BASIC and wrote programs on paper, working out in his head what the machine would do with each line. With nothing to catch his mistakes, he had to understand the whole program before it could ever run. That habit of thinking a system through before running it has stayed with him ever since. He trained in advanced engineering and NC/CNC programming at SOUaU Svitavy and graduated with honors in 1998. Instead of going to university, he went to work. At Aerotechnik and Metra, he wrote programs for CNC machines used in aerospace and manufacturing projects. He then moved into systems engineering and Oracle databases for municipal government agencies, and in 2009 he joined SYSCOM Software, where he would spend more than a decade. There he built financial software for the Czech Ministry of Finance, developed secure platforms for encrypted calls and messaging, and eventually became the company's Enterprise Architect. His first serious AI project came through a grant from the Technology Agency of the Czech Republic, which he won and led together with Faculty Hospital Vinohrady in Prague. His team of eight built a system that uses a convolutional neural network to recognize and classify burn injuries from images. Medical imaging left no room for shortcuts, so Volek had to bring model training and continuous learning up to clinical standards. Alongside that work, he led engineering teams as a contractor for Packeta Group and InterCars Deutschland, and later for Foxconn, where 15 engineers under his direction built a Manufacturing Execution System. ## Measuring where AI compute should live Miroslav Volek began working with Personal AI as a consultant in the winter of 2022 and became its Head of Engineering in June 2025. Volek owns its engineering organization and technical strategy, from small language models and voice AI to the infrastructure that serves them, and he leads the company's technical work with major U.S. telecom operators, meeting their teams every week to scope applications and carry designs through to delivery. Real-time voice AI turned an old infrastructure debate into a practical problem. A voice assistant has to answer within milliseconds, and every answer costs compute. An operator can pool its GPUs in one central data center or spread them across many smaller sites near its customers, and the usual rule of thumb says pooling is more efficient. Volek and his partners at a major U.S. telecom operator decided to test that assumption formally. The result is the co-authored paper *"Traffic Correlation and the Erosion of Pooling Benefit,"* which proposes a correction to the Kingman–Whitt approximation, a classic formula for estimating queues and waiting times. The paper shows how the advantage of pooling shrinks when requests arrive in correlated bursts, and it gives operators a way to choose between centralized and distributed compute using measurements from their own networks. The same approach shaped the distributed inference architecture Volek designed for Personal AI across NVIDIA GPUs and AWS infrastructure. It now supports Personal AI's role in [Comcast's AI Grid edge field trial](https://www.personal.ai/pi-ai/personal-ai-at-nvidia-gtc-bringing-memory-to-the-ai-grid), announced at GTC, where the company's small language model and memory platform runs the Small Business Concierge Agent, one of three launch use cases. Volek's work also reaches into the network itself. As part of the AWS–NVIDIA Telco AI Fellowship, which selected six startups worldwide, he created an architecture that embeds lightweight AI models into the radio layer of mobile networks, so that a network can configure and optimize itself. ## Toward networks with AI built in Miroslav Volek sees these projects as steps toward making AI a native part of network infrastructure. In the near term, he plans to help Personal AI take its work with U.S. telecom operators from field trials to national deployment, alongside the company's partnerships with SoftBank and NVIDIA. Further ahead, he wants to contribute to the technical foundation of AI-native 6G networks, which are expected to support uses ranging from autonomous vehicles to connected medical devices. His approach to getting there has not changed since the days of writing BASIC on paper. *"An idea isn't finished until it runs reliably on real hardware for real users,"* Volek says. The distributed inference work followed that rule from the beginning: it started as a practical question about where compute should sit and ended as a published framework that any operator can test against its own data. The boy who once watched a joystick move a figure across his cousin's screen now designs systems in which a person speaks and the network answers in less than half a second. © Copyright IBTimes 2026. All rights reserved.