Projects
The workspace is available as a static view. Read the work below or use the CV and contact links above.
OncoLA
Turning a patient’s everyday text message into information a care team can use.
Semifinalist, Ochsner Health / ASCO Healthcare Hackathon. Draft SOAP notes require clinician approval.
Project notes
Patients and caregivers report chemotherapy symptoms through SMS. A Groq-hosted open-weight gpt-oss-120b model structures free-form English, French, and Spanish messages.
The daily symptom-risk classifier achieved 0.78 recall and 0.68 precision on held-out data. These measurements describe that classifier, not clinical effectiveness. A separate forecaster estimates seven-day hospitalization risk; caregiver burden is tracked separately.
Built with Next.js, TypeScript, Prisma, and SQLite. Verified through 143 automated tests, a concurrency load test, and FHIR-lite export validation against an HL7 reference validator.
LIDC Reader Disagreement
When radiologists disagree, can a model’s uncertainty help identify the difficult cases?
Predeclared analyses, blinded review, and patient-level split validation keep the investigation accountable.
Project notes
The study investigates whether CT-model uncertainty predicts disagreement among LIDC-IDRI radiologists on noduleness. It is a research question, not a demonstrated finding.
A capacity-constrained max-flow algorithm reconciles overlapping annotations into candidate lesions. A blinded-review interface supports assessment without exposing the signals being investigated.
A 90-test synthetic suite validates patient-level split logic. The preregistered protocol and predeclared analyses preserve research integrity; interpretation must account for annotation reconciliation and the limits of this dataset.
Lung Nodule Classification
A three-branch model that makes room for uncertainty alongside a lung-nodule classification.
A dedicated uncertainty-estimation branch complements a three-branch ResNet-18 architecture.
Project notes
Uses public LUNA16 CT data. The preprocessing pipeline converts DICOM series and LUNA16 coordinates and diameters into model-ready inputs.
No performance or clinical-validation claim is presented here. Evaluation details will accompany the complete case study.
RIFT-NYC
Bringing physical constraints into models of pipe stress and mechanical systems.
PDE-based and variational methods, developed through cross-functional collaboration.
Project notes
Contributed modular backend architecture, data-generation and preprocessing pipelines, and TensorBoard experiment tracking. Model outputs were validated against mechanical-engineering baselines.
Career
SQL & data systems
Aug 2020 – May 2022 · Nawaratna EBS
Migrated paper records into SQL, built reporting and Python dashboards, and used OpenCV for document processing.
Mathematics Tutor → Senior Mathematics Tutor
Sep 2023 – May 2026 · Southeastern Louisiana University
Tutored Statistics, Calculus I–III, Linear Algebra, Differential Equations, and Numerical Methods, translating difficult concepts into understandable steps. Advanced to Senior Mathematics Tutor, mentored incoming tutors, and developed study guides used across course sections.
Applied AI
Jun – Aug 2025 · Nawaratna Life School · AI/ML Intern
Built a retrieval-augmented chatbot and an AI-writing triage classifier with privacy safeguards and human review. Benchmarked precision and recall.
RIFT-NYC
Jul 2025 – Feb 2026 · Physics-informed modeling
PDE-based and variational methods, modular pipelines, and comparisons with mechanical-engineering baselines.
Lung Nodule Classification
Feb – Jul 2026 · Medical machine learning
Three-branch ResNet-18 and CT preprocessing, with a dedicated uncertainty-estimation branch.
Software Developer
Jun 2026 – Present · SELU · English Department
Developing the Early Ruskin Manuscripts archive: TEI-XML publishing, PHP/MySQL, Dockerized TLS-secured Elasticsearch, and historical visualizations.
OncoLA
Jul – Aug 2026 · Patient-monitoring application
Multilingual symptom intake and clinician-approved note drafts. Semifinalist in the Ochsner Health / ASCO Healthcare Hackathon.
LIDC Reader Disagreement
Aug 2026 · Preregistered research study
Investigates uncertainty and reader disagreement using predeclared analyses, annotation reconciliation, and blinded review.
B.S. Computer Science
May 2027 · expected · Southeastern Louisiana University
Computer Science student focused on applied AI/ML, data systems, and trustworthy model behavior.
