We have long researched AI system design and deployment, working on both experimental and production-ready systems in domains like healthcare, sports, and data analytics. We develop solutions spanning from raw data ingestion to real-time AI inference and decision support, actively tested on real datasets and users.
However, there are many open challenges to address, such as:
- AI workload optimization (how to run large models efficiently, reduce inference latency, and minimize resource usage without sacrificing performance)
- Data processing pipelines (designing robust ETL (Extract, Transform, Load) pipelines for heterogeneous data sources, supporting both batch and streaming workloads)
- Scheduling and orchestration (deciding how and where to run different AI tasks, from GPUs to edge devices, to achieve the best performance-cost tradeoff)
- Model serving and scaling (deploying models that adapt to varying workloads, ensuring reliable performance under peak demand)
- Data storage and retrieval (optimizing large dataset indexing, caching strategies, and query execution for fast access)
- Leverage open-source LLMs (Look at using open models like LLaMA or the new OpenAI OSS models for domain-specific applications while managing deployment complexity, performance tuning, and integration challenges)
- Monitoring and fault tolerance (detecting bottlenecks, diagnosing failures, and maintaining high availability for AI services)
- Security and privacy-aware processing (ensuring sensitive data is processed securely, including encryption and compliance with regulations)
Want to be part of our research group and improve the performance and reliability of AI systems? Contact us and we’ll find a thesis topic that matches your interests.
Goal
To design, implement, and evaluate methods for improving AI system performance, scalability, and efficiency across data processing and model execution pipelines. Understand the tradeoffs between an AI system’s performance (processing costs, latencies, resource consumption, etc.) versus the quality of the output.
Learning outcome
- Understanding of AI workload characteristics and optimization techniques
- Experience with large-scale data processing frameworks
- Knowledge of deployment strategies for AI models
- Skills in system performance analysis and tuning
- Exposure to real-world datasets and use cases
Qualifications
Strong motivation, basic programming experience, interest in AI systems and performance optimization
Contacts:
Pål Halvorsen <paalh@ifi.uio.no>, <paalh@simula.no>
Steven Hicks <steven@simula.no>