I'm interested in systems that generate or adapt something for a person, and in what happens when the system gets it wrong. So far that has meant music visualization for relaxation, AI-drafted clinical letters that a doctor has to sign off on, and testing how medical LLMs handle a patient whose condition keeps changing.
We tested a soft and an intense music-driven visualization with 30 participants, using PSS-10, PANAS, an aesthetic-pleasure scale and interviews. Both versions lowered negative affect in the short term, the soft one more so, and people who found the visuals more pleasing also reported more stress relief. It was a small group of young adults in one session, so it says nothing about long-term effects.
I built the TouchDesigner and Python system, contributed to the study, and wrote the mathematical description of the audio preprocessing and the audio-to-visual mappings for the paper. More on the project page.
Polaris drafts referral letters with a language model. I built the generation pipeline and the review flow around it. The model only sees the clinical context the patient agreed to share, its output is checked against a schema and its citations, and a rule-based draft takes over when generation fails. Nothing is sent until a doctor edits, approves or rejects it, and every decision is kept.
We are building a clinical benchmark on Synthea where the simulated patient's state keeps changing and complications can come up, so a model has to keep making decisions as the case unfolds. I work on the workflow that runs several LLMs on the same cases and records what each one does, and on metrics for repeated orders, delayed treatment and how complications are handled. When an LLM judge scores a run, we check its assessment against the encounter records.
The follow-up to the published study. I cleaned and time-aligned EEG and PPG recordings with the audio and visual stimuli, helped set up an offline RL framework that uses those signals to adjust the visuals, and compared BCQ, IQL and CQL for learning from recorded data. I also drafted the motivation and related-work sections.
If any of this overlaps with what you work on, I'd like to hear from you at [email protected].