I’m a master’s student in Applied Mathematics at Aalto University with a background in Engineering Physics. I’m broadly interested in machine learning, and my recent work has focused on reinforcement learning and language model training.
- RL scaling on LUMI supercomputer — Studied weak and strong scaling of Qwen3-8B reinforcement learning workloads across 1–16 LUMI GPU nodes, measuring throughput and analysing performance bottlenecks. Includes setup and launch scripts and a technical article.
- RL for LLM agents — Compared six reinforcement learning algorithms for multi-turn agents in the ALFWorld and WebShop environments, analysing task success, training dynamics, and computational cost. Includes a report and launch scripts. Added GSPO support to the verl-agent training framework.
- Active learning for molecular simulations — Developed a workflow to train MACE models for molecular dynamics simulations. The workflow trains a group of models, identifies configurations where their predictions disagree, and uses quantum-mechanical calculations to label those configurations before adding them to the training data and retraining the models.
- ML library from scratch — Implemented regression models, SVMs, and neural networks in NumPy. A CNN built with the library achieved 98.13% accuracy on an MNIST test split for handwritten digit recognition.
- Character-level Transformer — Implemented and trained a decoder-only Transformer in PyTorch on the first Harry Potter book to generate text character by character.