Virtual Human
Future Software Tech
Semester programme:Software Engineering
Ben van der Linden
Junjie Cen
Jonathan Christyadi
Jak Benev
Viktoria Todorova
John Ooi
Project description
The Virtual Human project aims to create an AI-driven digital avatar capable of real-time interaction through natural language and expressive behavior. The system integrates AI models with containerized services and is deployed in a Kubernetes environment for scalability and reliability. A central focus of the project is improving deployment efficiency, monitoring, and system modularity.
Main research question:
What are the challenges for transitioning to a Kubernetes environment, packaging and deploying AI models, and ensuring real-time monitoring and low-latency interactions in virtual human systems?
Context
The Virtual Human project is part of the “Digital Innovation District for Society” initiative, supported by MindLabs and Fontys. This project aims to deploy AI-driven virtual agents in scalable and cloud-native environments. A major challenge is to automate the deployment of configurable AI components. Using YAML-based configurations, a custom OCI-compliant service builds and deploys model-specific Docker containers to a Kubernetes cluster. This modular design allows dynamic scaling, real-time monitoring, and better control of virtual human behavior.
Results
The main outcomes of this project are a working YAML-based deployment system and an OCI-compliant container builder that can automatically deploy AI models as virtual humans in a Kubernetes environment. These tools allow users to define model settings in a single YAML file, which is then parsed into build instructions and used to create and launch Docker containers. The containers are equipped to run AI models with preloaded dependencies, optimized startup, and real-time API endpoints. For now, the system uses pre-trained models from Hugging Face.
Validation was done through successful local deployments using Minikube and a live deployment on Netlab. Both setups showed consistent performance, model availability, and fast container boot times. Error handling was also built in, allowing failed models to be skipped without stopping the pipeline.
Based on the Technology Readiness Level (TRL) scale, the project reached TRL 4, demonstrating that the concept and core components work in a controlled setting. These results show the system’s value in automating virtual human deployment and provide a strong foundation for future scalability, GPU optimization, and real-world use in digital interaction platforms.