Research

This began in the CRA UR2PhD research training program (June to July 2026). There I built a noise-controlled evaluation harness that separates genuinely decodable neural signal from what a language model would produce anyway. It came with probing and retrieval baselines and a CLIP-style contrastive EEG encoder, packaged as a tested Python library of roughly 2,500 lines. I also surveyed privacy-preserving machine learning for neural data and wrote the program’s research proposal on a cross-subject, federated decoding direction.

At the SPEN Lab I am continuing this work with reproducible signal-processing, training, and evaluation pipelines for cross-domain EEG experiments. There are no publications yet.

I led a three-person project that simulates event-triggered recording on the ESC-50 audio corpus, with Poisson event arrivals, 10 dB signal-to-noise ratio and 30% overlap. Features are MFCC, delta-MFCC and RMS, and the classifier is a Random Forest.

Under realistic streaming, the proposed adaptive policy captured 77.6% of events, 31.8 points more than a random-trigger baseline with matched storage (Cohen’s d = 2.59, p = 0.0004, bootstrapped over 10 runs), while cutting stored data by 87%.

This is a simulation study. It was not deployed on hardware.

I joined this collaboration in Fall 2026. My own contribution to the next phase is still being defined, so there is nothing to report yet.