Hack & Tell: Music Gen, Real-Time Avatars, Vectorial AI & Multi-Agent Systems

Join us for a very special Hack & Tell hosted by Allianz Technology!

We’ve curated a lineup of four talks spread across 2 tracks that dive into the intersection of high-end research and creative engineering - plus plenty of time to meet other AI engineers, researchers and founders over pizza and drinks.

This time, we’ll be looking at how AI is built and deployed across four distinct domains:

🎵 Training AI on personal musical styles
🖊️ Building editable vectorial design systems
👤 The math behind real-time avatars
⚙️ Managing context within large-scale agentic workflows

---------------------------------
🗓 Agenda
---------------------------------

18:00 – Doors open (Arrive early for security)
18:30 – Start of Talks (You can switch tracks between talks)

Track 1:

• Teaching AI to Make Music Like I Do
• Building a Real-Time Interactive Avatar Pipeline

Track 2:

• Penpot Wizard: Bridging Design and AI
• Managing Context Bloat in Agentic Workflows

20:00 – Networking (Pizza & Drinks)
21:00 – End

---------------------------------
📍 Location
---------------------------------

Allianz Technology, Av. d'Icària, 199, Sant Martí, 08005 Barcelona

⚠️ Important: Registration with DNI/NIE is required (*)

---------------------------------
🎵👤 Track 1
---------------------------------

🎤 Bianca Cheng – Teaching AI to Make Music Like I Do
Bianca is a musician, designer, and technologist (formerly at Apple and MIT Media Lab) who approaches engineering with a creative soul. She’ll be sharing her work with HeartMula, a project where she trained models on recordings of her own songs. Learn how she uses AI as a collaborative partner to explore new musical ideas while staying true to her unique creative voice.

🎤 Arnab Dey & Marcel Alcoverro (UNITH) – Building a Real-Time Interactive Avatar Pipeline
What does it actually take to turn an image into a real-time photorealistic human avatar? Arnab (Senior Research Engineer) and Marcel (Head of AI) will walk through the full engineering stack behind a live interactive avatar: from LLM-driven speech synthesis and autoregressive 3D motion generation to Gaussian Splatting rendering streamed at 60 FPS. This talk dives into the research models powering visual avatars, the hard latency trade-offs that shape every design decision, and what production MLOps looks like when your service is a human face.

---------------------------------
🖊️⚙️ Track 2
---------------------------------

🎤 Xesca Alabart (Gluecharm) – Penpot Wizard: Bridging Design and AI
Why do most AI design tools give you code instead of editable design objects? Current tools like v0 or Bolt often leave designers stuck in a "prompt loop" - where even a minor layout tweak requires a new prompt rather than a simple click-and-drag.
Xesca, creator of Penpot Wizard, is pioneering a fundamentally different path. Her talk introduces a Vectorial AI-driven methodology that treats the canvas as a live environment rather than a code-generation target. By designing with AI and Vectorial Design as a structural foundation, Penpot Wizard generates native design objects—shapes, paths, and components—that designers can edit normally. It’s about moving past the prompt-only bottleneck to provide a functional, editable starting point for real-world UX.

🎤 Gabriel Llauradó Queralt (Allianz Technology) - Investigator: A Multi-Agent System That Digs into Tickets
Every Jira ticket hides a trail across code, databases, logs, and half-forgotten Slack decisions. Gabriel built Investigator using LangGraph, with specialist agents for code, DB, Jira, OpenSearch, Confluence, GitHub and Vision — Orchestrated by a Planner, Router, Critic and Evaluator that argue with each other until the evidence holds up. He'll walk through the architecture, the prompts that actually work, human-in-the-loop checkpoints, and lessons learned making LLMs reason about real enterprise codebases.

See you soon!

(*) You will greatly help us by providing your DNI/NIE upfront when you RSVP. It is solely used for access to the building and will not be used or stored for any other purpose. In any case, you can also answer with "I want to register onsite". In that case, please show up early at 18:00 to guarantee that the registration can be finished before the event starts.