How can models learn continuously from streaming data without forgetting what they’ve learned? My thesis at Scania explores this directly — online learning algorithms that adapt to changing vehicle behavior over time while running on edge hardware. Also connected to federated settings where fleet-wide models must incorporate new knowledge from individual trucks without degrading.

Beyond correlation: understanding cause and effect in data. Interested in how causal structure can improve anomaly detection — in vehicle sensor data, anomalies often manifest as breakdowns in physical relationships between sensors, not just unusual values. Also exploring causal inference for interpretable AI systems.

Rethinking how models learn. Kore explores this — a language where the compiler acts as a reward function for LLM code generation, replacing the noisy “run and hope” loop with deterministic verification at compile time. More broadly interested in energy-based models, biological plausibility, and optimization approaches beyond standard backpropagation.