Financial Virtual Assistant for Portfolio Management

• Name: Financial Virtual Assistant for Portfolio Management , FVAP
• EuroHPC machine used: MeluXina
• Topic: Computer and information sciences

Overview of the project

FVAP aims to develop a Turkish large language model specialised in investment funds, portfolio management, and financial advisory services. The system is designed to understand user queries, extract intents and entities, determine investor risk profiles, and generate context-aware portfolio allocation recommendations. The project combines domain-specific NLP, large-scale model training, and distributed GPU computing to build a scalable virtual assistant that can later be adapted to different financial institutions.

How did EPICURE support the project and what were the benefits of the support?

“The project required efficient execution of large-scale LLM training and fine-tuning workloads on EuroHPC systems using multi-GPU and multi-node configurations. During the process, we encountered several technical challenges, including CUDA–PyTorch compatibility issues, inefficient GPU utilisation, communication overheads, and performance bottlenecks — particularly those related to the BitsandBytes library.

The EPICURE application support team assisted us in configuring and validating the software environment, selecting appropriate library and driver versions, and adapting the training workflows to the architecture of the target system. They supported benchmarking activities, analysed scaling behaviour, and helped optimise communication patterns across GPUs. This guidance allowed us to stabilise distributed training runs and significantly reduce inefficiencies.

The support resulted in faster and more stable training cycles, improved parallel efficiency, and better utilisation of allocated computing resources. Considerable time was saved by resolving environment and compatibility problems early. The team was able to focus more on model quality and experimentation rather than infrastructure troubleshooting. The optimisations enabled larger-scale experiments and improved accuracy in intent detection and investor profiling tasks.” – Derya Uysal