MIA — Generative AI hyperrealistic avatar
- Role
- Frontend Tech Lead
- Year
- 2024 — 2025
- Stack
- ReactTypeScriptAzure OpenAIWebGL
01 /Context
W2M, the technology arm of Grupo Iberostar, wanted to introduce its tourism distribution platform at FITUR with something the industry had not seen before: a hyperrealistic digital person able to hold real conversations on stage. The brief asked for an interactive avatar that could understand and respond in natural language, sustain a credible presence in front of press and partners, and serve as a public showcase of what generative AI looks like applied to travel.
02 /My role
I led the frontend across both editions of the project. The first version focused on the conversational layer — wiring the React interface to Azure OpenAI for reasoning, Avatar Custom for the rendered persona, and a neural voice profile for the spoken response. The second version, for FITUR 2025, added a Holobox 3D installation, which forced a rework of the rendering pipeline and the synchronization between speech, lip-sync and ambient feedback. I owned the architecture decisions, the integration boundaries with the AI services team, and the end-to-end behaviour shown on stage.
03 /Technical highlights
- React + TypeScript front controlling the avatar lifecycle: idle, listening, thinking, speaking.
- Integration with Azure OpenAI for reasoning and a custom neural voice for the spoken output.
- Avatar Custom rendering pipeline tuned for low-latency response under live stage conditions.
- Second version reworked to drive a Holobox 3D installation, with new synchronization between voice and 3D output.
- Stage-grade resilience: graceful fallback when the network, the model or the rendering surface degraded mid-demo.
04 /Outcome
MIA was the first generative AI hyperrealistic digital person presented at FITUR. The project was covered by CIO España, Microsoft News and El Español, and returned the following year in a more ambitious Holobox 3D format. Beyond the press moment, it became Bravent's reference case for what an enterprise-grade GenAI experience looks like when the frontend has to hold up live in front of an audience.