Reproducible scientific software, and a couple of shipped full-stack apps — the engineering behind the research.
End-to-end, reproducible pipeline mapping language onto single-neuron activity: 57-feature / 9-layer linguistic extraction → adversarially purified GCN + SBERT embeddings → cross-validated Poisson / ridge encoding with confound controls. Scales to 435 neurons across 7,346 word-level timepoints.
A question-answer embedding framework that extracts interpretable structure from LLM hidden states and aligns it to neural population geometry via RDM/RSA and Procrustes, with length-controlled brain-score and wavelet null models.
Poisson GLM suite with nested likelihood-ratio tests, circular pitch encoding, MFCC spectral PCs, and multi-system clock-drift compensation — plus a UMAP-based spike-sorting quality-triage tool for waveform clustering diagnostics.
An interactive web map for discovering Houston restaurants scraped from Instagram, with location filtering, marker clustering, and a CSV→geocode→JSON data pipeline. A shipped, full-stack app — concrete evidence of production engineering.
A B2B concept for accurate pronunciation of names (especially non-English) at graduation ceremonies. Dual-mode: model-generated pronunciation via multilingual grapheme-to-phoneme + a curated lexicon (the competitive wedge), plus self-recording capture. IPA as source of truth; all voice data self-hosted (FERPA-compliant).