Neuroscience researcher and biomedical engineer.

Understanding nociception across brain networks

For my PhD, I design and implement the technical workflow for studying neural responses across S1, PFC, and ACC-targeted recordings. The work spans data organisation and quality control, multiregion physiology, time–frequency analysis, explainable CNNs, and animal-level transformer benchmarking.

I work across the experiment, data pipeline, physiological analysis, and model evaluation while retaining provenance at each stage. Because the recordings are made under general anaesthesia, I keep the interpretation to stimulation-related and nociception-related activity—not conscious pain or a clinically validated biomarker.

Spinal cord stimulation modelling

During a visiting period at Duke's Grill Lab, I turned a neuromodulation question into a Python simulation study. Using PyFibers and an adapted McIntyre–Richardson–Grill axon model, I implemented bipolar stimulation, compared conventional, FAST, and burst waveforms across fibre diameters and charge-balancing strategies, and built threshold and fidelity sweeps that led to an IEEE EMBC paper.

The analysis separated waveform-timing effects from pulse-width effects: burst stimulation had lower thresholds under the tested settings but less consistent pulse following. The single-fibre model tests candidate mechanisms rather than making a direct clinical prediction.

TDT-to-BIDS electrophysiology pipeline

I designed and built the data layer behind the PhD analyses. It converts raw recordings from a Tucker-Davis Technologies RZ2, RS4 storage, and Synapse into structured BIDS-FIF data, resolves simultaneous S1/PFC and ACC streams, reconstructs stimulation events, and produces analysis-ready epochs.

Rather than hiding manual decisions, the workflow records acquisition pairing, source and configuration fingerprints, channel and epoch quality control, and review status. This makes downstream analyses easier to trace and rerun across animals and recording sessions.

Trading-card identification and valuation

As a personal project, I built a mobile-first retrieval tool for identifying Japanese trading cards from shop photos. It combines DINOv3 embeddings, CUDA-accelerated cosine search, PaddleOCR, market data, and a Next.js interface. I chose retrieval so new cards can be added by rebuilding the local index rather than retraining a classifier.