The person behind the platform
Jesper Rietbergen
Founder and CEO, Flireo B.V.
I taught a lot of people how to build voice agents on someone else's platform. Then I built the layer underneath, because that is where the calls actually break.
I run Flireo B.V. from Deventer in the Netherlands, and I am the person who writes the orchestration code that VoiceDock runs on. Not a founder who used to be technical. When a customer's telephony fails on a Tuesday evening, I am the one reading the SIP trunk logs, and everything published on this site is written from that seat.
How I got here
Four wrong turns that added up to the right one
None of this was a plan. It was a sequence of things that stopped working, and each one taught me something specific about where the durable position sits.
- 01
2023: a chatbot that a general model made pointless
I started Flireo in November 2023 on a simple bet: AI was going to matter, and the way to have a chance was to specialise from day one. Our first product was a chatbot answering occupational-health and safety questions against a curated knowledge base. Within a year the general models could answer those questions better than we could, and the reason for it to exist evaporated. That is the first lesson I actually paid for: a thin layer over somebody else's model is not a company.
- 02
2024: finding voice, and finding an audience
I went looking for something harder and landed on voice. I taught myself Vapi, the APIs around it, and the automation glue, and started publishing what I learned on YouTube. Not clips. One to three hour walkthroughs where you build an agent end to end, in real time, including the parts that go wrong. The channel became Flireo's growth engine before we had a sales process at all, and most of the relationships this company runs on started there.
- 03
The two side businesses I stopped on purpose
Two things sold well and I shut both down. Consultancy at 125 euro an hour, one or two hours of live troubleshooting, and prebuilt automation templates at 625 euro each. The templates are the interesting failure: people bought the outcome without wanting the understanding, then came back frustrated by problems the template was never going to solve. Selling ready-made to people who did not want to learn the system produced unhappy customers on both sides. That conclusion is where done-with-you comes from. It is not a marketing angle, it is what was left after the alternative failed in front of me.
- 04
Down the stack, one ceiling at a time
The tooling moved as the problems got harder. No-code front ends, then a backend-as-a-service, then a low-code automation engine, and each move happened because I hit the ceiling of the layer above it. The last ceiling was the important one: when a call drops and the platform tells you it failed, and you have no way to look any deeper, you are stuck holding a customer problem you cannot fix. So I stopped renting the layer and started writing it.
What I do now
I build it, I run it, and I pick up the phone
VoiceDock is a European voice-AI orchestration platform, and I am the one maintaining the part of it that other platforms treat as a black box.
Python and LiveKit for the orchestrator, Next.js for the dashboard and this site, Hono and Node for the services, Supabase for data, Hetzner and DigitalOcean for compute, Telnyx for telephony. European infrastructure, on purpose.
Writing
What I publish here
Everything below is written from production experience rather than from a content brief. If a number appears in one of these pieces, it is either measured, cited, or explicitly labelled as illustrative.
The Latency Budget of a Phone Call
Every platform quotes one latency number. A caller experiences a chain of eight stages, only two of which are the model. Where the milliseconds actually go, which ones are physics, and why turn detection usually costs more than time to first token.
9 min readBarge-in That Actually Works: Tuning VAD for Phone Audio
Voice activity detection settings tuned on clean studio audio fall apart on an 8 kHz phone line. The two jobs one detector is asked to do, the false barge-in and how to recover from it, and why we moved turn detection back out of the cloud.
10 min readEU AI Act Article 50: What Voice Agents Must Disclose
From 2 August 2026 a caller must know they are talking to a machine. What Article 50 actually requires of a phone agent, who carries the obligation, and why proving it is a job for the orchestration layer.
11 min readWhat a Voice AI Agent Really Costs Per Minute
Platform fees are the small half of the bill. A component-by-component breakdown of what a minute of AI phone conversation costs, where the hidden multipliers sit, and how to model it before you commit.
12 min readWhen the Transfer Fails: Handing a Call to a Human
Every voice agent demo transfers to a human on the first try. Production is not a demo. The failure modes of a warm transfer, why silent fallbacks are worse than a refusal, and what to build instead.
10 min readWhy Voice Agent Calls Fail: Reading the SIP Logs
A dashboard that says 'error' tells you nothing. The four layers a phone call fails at, the SIP response codes and Q.850 causes that name the cause, and a triage order that converges instead of guessing.
13 min readSelf-Hosted Voice AI: Who Holds the Pager?
Self-hosting a voice stack to escape per-minute pricing is a real strategy with a real bill attached, and the bill is mostly operational. What you take on the day you own the pager, and the three honest ways out.
12 min readGemini Live 3.1 Prompting Guide (Voice Agents)
You can't prompt Gemini 3.1 Flash Live like a text LLM or a TTS engine. A practical prompting guide for native-audio voice agents: instruction design, tone, language, and what quietly fails.
9 min readElsewhere
Where else to find me
The long-form video work is where most people met me first, and it is still the most honest look at how I actually work.
Want to talk to the person who builds it?
There is no sales layer between you and the engineering here. If you have a voice problem worth solving, you will be talking to me.