Research
Peer-reviewed work at the intersection of deep learning and applied systems: Document AI, privacy, clinical summarization, and parameter-efficient fine-tuning.
3 publications
- EMNLP 2024 (Main)
De-Identification of Sensitive Personal Data in IIT-CDIP-Derived Datasets
A pipeline for detecting and redacting sensitive personal data in large document-image corpora derived from IIT-CDIP, enabling privacy-preserving research on real-world Document AI at scale.
nlpdocument-aiprivacyread paper ↗ - Elsevier — Applied Soft Computing
Parameter-Efficient Fine-Tuning for Hospital Discharge Summarization
PEFT methods applied to clinical summarization, matching full fine-tuning quality on hospital discharge summaries at a fraction of the trainable parameters and compute.
peftllmhealthcareread paper ↗ - Elsevier — Applied Soft Computing
DoRA+: Enhancing Weight-Decomposed Low-Rank Adaptation
An extension to weight-decomposed low-rank adaptation (DoRA) that improves the magnitude/direction decomposition for more effective and stable parameter-efficient fine-tuning.
lorapeftdeep-learningread paper ↗
Interested in the implementation?
I write about what happens after the paper: evaluation, reliability, latency, cost, and operating the system in production.