Time-series language models
Aug 2026 –
Independent research
The direction came out of the ETH Agentic Systems Lab selection, where I reached the deep-dive interview; the work itself is independent and not an ETH project. It builds on the lab's TimeRLM training pipeline and AnomalyXL benchmark. I replaced the encoder with a multi-resolution Haar-wavelet encoder whose output goes into Qwen3.5-4B as soft tokens (LoRA), and added a zoom action that lets the model re-read a window of the signal it chooses.
On two held-out confirmation sets, zoom plus two small fitted corrections for the window's centre and length localized 112/157 and 113/161 anomalies at IoU ≥ 0.5, against 88 and 77 for the reference model. The zoom on its own did not clear that bar, and GRPO post-training gave no real gain. Integrity checks along the way found and quarantined a train/test leak and led me to retract an earlier result that came from a coordinate shortcut.
Airway analysis in chest CT
May 2026 –
Remote collaboration with researchers at the University of Michigan and Emory University
The agreed starting task was estimating the airway wall around a given lumen segmentation, on the public ATM'22 data. On my own initiative I prototyped two machine-learning extensions. The first repairs breaks in segmented airway trees with Monte Carlo Tree Search; a controlled experiment showed that, with that reward, search chose the same links as a greedy rule. The second trains "judges" that decide whether a gap in an nnU-Net mask is a real missing segment (a multimodal LLM with a 3D wavelet encoder, a small probe, gradient boosting), with nnU-Net baselines on the Politecnico HPC cluster. The judges worked on synthetic breaks but did not transfer to real ones, so I archived that line.
Simulation Engineer, Rilemo
Jul – Sep 2026
Medtech startup, remote
Rilemo builds a portable electromagnetic imaging device for the head. I worked on electromagnetic simulation and on generating 3D multi-tissue head models for data augmentation. Details are under NDA.