Nova Replica

At Nova Replica, we worked on interactive AI avatars capable of holding real-time video conversations.

I co-founded the project and focused primarily on the engineering behind running and orchestrating the AI models that made the experience possible.

Making research models work in practice

One of the biggest challenges was the gap between a research demonstration and a system that could actually run reliably.

We experimented with models from research papers covering different parts of the avatar pipeline, including speech and visual generation.

Although individual models could demonstrate real-time performance, their reference implementations weren’t necessarily designed for efficient deployment. Loading models, managing memory, coordinating inference, and keeping the entire pipeline responsive introduced a different set of problems.

Much of my work involved digging into model implementations and optimizing how they ran together.

I worked on parallel execution, memory management, and Python event-loop behavior to reduce unnecessary overhead and coordinate the different stages of inference.

I also adapted open-source English text-to-speech models for Turkish and built a system orchestrating four AI models to support interactive video conversations.

Beyond the prototype

The resulting system supported real-time AI avatar conversations, and we worked on projects involving Garanti BBVA and Müzeverse.

Nova Replica taught me something that has stuck with me: getting a research model to run is one challenge. Optimizing it, designing the infrastructure around it, and making it production-ready is an entirely different one.