Experience

• Engineered a data-cleaning pipeline for noisy, multilingual field audio.
• Built a detection pipeline that combines two pretrained laughter models, Google’s YAMNet and Gillick et al.’s ResNet, and ran it across all 1.6M files in the company’s audio database to mine clips for a training and test set.
• Fine-tuned the Gillick model in PyTorch on that set, reaching 94% accuracy.
• Presented progress to the CTO and CEO.
• The feature is now being shipped to FactorLab’s platform.

TDA framework for neurodegenerative drug discovery

• Profiled 345 FDA-approved and clinical-stage compounds across 7 neurodegenerative diseases.
• Built feature sets from 94 physicochemical descriptors and 881-bit PubChem fingerprints, pulled from 5 public databases.
• Ran six UMAP/t-SNE Mapper configurations and disease-specific persistent homology.
• Found CYP450 metabolic interactions driving cross-disease structure in 4 of 7 diseases.
• Built a reusable chemical space for projecting new compounds, validated with leave-one-out testing.

PTSD treatment meta-analysis

• Led a 5-person team analyzing 538 effect sizes from 328 RCTs.
• Applied three-level meta-regression with robust variance estimation, cross-checked by a MetaForest ML screen.
• Showed comparator type drives effect size: g = −0.64 vs. inactive controls, near zero head-to-head.
• Re-estimated the field-wide effect with PET–PEESE, selection models, and RoBMA, which put it near zero after bias correction.

Coronavirus host-network study (in progress)

• Leading a team mapping SARS-CoV-1, SARS-CoV-2, and MERS host proteins onto the human interactome (BioGRID, STRING).
• Using persistent homology to compare the network topology each virus targets.

• Fine-tuned 15 pretrained transformer models (BERT, RoBERTa, ALBERT, DistilBERT, MiniLM, and others) for 6-class emotion classification on the 20K-sample dair-ai/emotion dataset.
• Built the training and evaluation pipeline with the Hugging Face Transformers and Trainer APIs, scoring each model on accuracy, precision, recall, and F1.
• Correlated each model’s architecture and pretraining features with its performance. The number of attention heads predicted better results (r ≈ 0.5), while more layers predicted worse ones (r ≈ −0.8).
• Top model: RoBERTa-base, with 93.0% accuracy and 93.0% F1.
• Applied the findings to an emotion-driven book recommendation framework, which became the basis for MoodReads.
• Published in the Journal of Emerging Investigators (April 2026) as “Evaluating key factors in emotion detection models for AI-driven personalized bibliotherapy”.