September 3, 2026
Comments Off on AI Tinkerers Madrid: AI in Real-Time Communication Systems

Now that everyone’s back from summer break here in Madrid, it feels like the right time to recap AI Tinkerers Madrid: AI in Real-Time Communication Systems, which WebRTC.ventures sponsored and hosted this past July alongside BTS, who provided their offices and a great spread of local snacks. We aimed for 30 attendees and comfortably surpassed that number. The feedback was strong, too! People are already asking about the next edition, and I’ll have news on that soon.

For those unfamiliar with the spirit of AI Tinkerers, it’s a global movement built around hands-on people using AI in their workflows or their products. The premise is simple: show, don’t tell. Slides are discouraged, code is always welcome, and a working demo beats both.

Since I work as Project Lead at WebRTC.ventures, where we build custom, production-ready real-time communication and AI systems across a range of industries, and BTS also works in communications, we focused this edition on AI applied to voice AI agents. That said, the mic was open to anyone who wanted to show their own industry or application, which is how we ended up with a genuinely varied lineup!

What we saw on stage

TalkingBits: real-time call translation (BTS)

Alejandro Pérez Malagón opened the night with TalkingBits, a voice AI agent that joins a call as a third participant on LiveKit and translates it live, in both directions. The problem statement on his slide summed it up well: two people are on a call and don’t share a language, so the agent sits in the room as an interpreter. Under the hood, the pipeline runs speech-to-text, translation through an LLM, and text-to-speech back into the call with a neat detail called “audio ducking,” which lowers the original speaker’s voice so it doesn’t compete with the translated audio. The agent can also dial out over a SIP trunk. Since it’s all built on LiveKit, a WebRTC platform, it was a great opening act for a crowd that cares about real-time infrastructure.

SHINE: real-time noise reduction under strict latency (BTS)

Aitor del Val Allueva followed with SHINE, BTS’s project for real-time noise reduction under strict latency limits. Rather than just describing the approach, the demo showed the system running: GPU inference server logs processing live audio, and a metrics dashboard tracking compute time and a real-time factor around 0.911: proof the model keeps up with the audio as it arrives rather than falling behind. Anyone who’s shipped real-time audio knows this tradeoff well: a noise suppression model that performs beautifully offline can fall apart the moment you need results in milliseconds. Seeing production numbers, not just a lab demo, made this one land.

A Pipecat-based behavior monitor (WebRTC.ventures)

Our own WebRTC Senior Engineer, Alfred Gonzalez, walked through a Pipecat pipeline built to catch inadequate behavior in live conversations. He was explicit about what it isn’t: not a turn-taking conversational agent, but a continuous evaluation monitor that runs alongside the live audio and video pipeline, classifying conversation frames against a set of defined risk categories using pluggable evaluation strategies. It was a good reminder that voice AI agents need guardrails to be production-ready, not just a working prototype. And, a solid showcase of the team’s hands-on experience with Pipecat, a framework that’s increasingly familiar to this audience.

Git-style branching for multi-agent conversations (WebRTC.ventures)

Jesús Leganés-Combarro closed our team’s contributions with a talk on ToonLogsTree. No live demo, but plenty of ideas to sit with. He opened with a joke slide, “Professional Cat Herder,” which set the tone for the real problem: managing multiple AI agents that need to collaborate is exactly that. His first principles were straightforward: agents are autonomous, they need to communicate, their conversations need to persist (sometimes in parallel), and you need an immutable, rollback-able history of what happened. As he put it, “Git already solves part of this persistence and versioning problem very well.” From there he introduced TOON, a compact format readable by both humans and LLMs that uses fewer tokens than JSON, and used it as both the persistence and communication format. The payoff was ToonLogsTree: what if a conversation could branch like a Git repository? A diagram showed a main conversation branching into technical, commercial, and academic sub-conversations, which is a clean, inspectable way to manage multi-agent handoffs and explore alternatives without losing the thread.

Zeus: chat-driven modeling that ships real microservices (Metadev)

Antonio Chamorro Manjón, from Metadev, closed the night with Zeus, a chat-driven modeling tool embedded in an architecture editor. He typed plain-language instructions: “create a model for a Microservice,” “add a nickname property to the Vendor class,” and watched the assistant update the formal model live. From there, Zeus generated real code from a catalog that included a .NET microservice and a Spring Boot Java service, among others. The demo didn’t stop at the diagram: he ran an actual dotnet publish, built a Docker image, and tested the running container live with a Postman-style API client.

What stood out

My takeaway: real-time communications is a well-established industry, with decades-old standards for call quality and response times. Seeing AI applied live to those same fundamentals pushes the boundaries of what’s possible, both functionally and for the end-user experience. Watching live translation happen without breaking the natural flow of conversation, and noise reduction working at scale in real time, was genuinely astonishing.

Recognizing undesired behavior in video calls stood out to me, too. Having worked on educational platforms for elementary-age children, I know how critical a tool like this is for keeping kids safe in real-time online environments.

I also hadn’t considered applying a Git-style branching model to formal conversation logs the way Jesús did. That’s engineering in its purest form: taking an existing tool and reconfiguring it for a new problem.

And then there’s Zeus. Anyone who’s watched an engineering team define a system, go back to the whiteboard, and rework it all over again will recognize the value of what this platform offers as it’s built to speed up exactly that cycle. Several attendees were engaged enough to ask if it worked outside the Microsoft ecosystem. It doesn’t yet, but that’s coming soon.

Why this matters for our clients

None of this was abstract for us. Real-time translation, noise suppression under latency constraints, guardrails for voice AI agents, structured multi-agent architectures, AI-assisted software design … these are the same categories of problems we solve for clients at WebRTC.ventures, whether we’re building a custom WebRTC application or engineering a voice AI agent. Watching other teams tackle these problems live is a good gut check: it confirms we’re building in the right direction, and it gives us fresh ideas to bring back to our own client work.

Thank you to BTS for hosting and for all the presenters representing such a range of applied AI work in one room. Events like this are exactly why we stay involved in the AI Tinkerers community: they keep us close to what other builders are solving, and they’re a great way to meet people working on the same hard problems we are. We’re already looking forward to the next edition.

If you’re working on a voice AI agent, a real-time communication product, or need to scale your engineering team to ship faster, we’d like to hear from you. Reach out to WebRTC.ventures to start the conversation.

Further Reading:

Recent Blog Posts