Research Software
Open-source research software supporting my work on coherent Raman imaging, model reliability, and biomedical image analysis. All code is available on GitHub.
The CARS/BCARS ecosystem
Machine-learning models for retrieving Raman information from CARS/BCARS spectra often perform well under the acquisition conditions they were trained on, and degrade under others. Because the ground-truth Raman signal is rarely available experimentally, this failure mode is hard to measure on real data.
These three packages address simulation, retrieval, and validation as one connected workflow:
Simulate → Retrieve → Validate
CARSBench: simulation and domain-generalization benchmark
Generates synthetic CARS/BCARS spectra together with their Raman-equivalent targets under controlled domain shifts, including spectral resolution, detector noise, baseline drift, calibration, spectral window, non-resonant background shape, and biochemical composition. Includes reproducible seeded generation, quality-control workflows, benchmark metrics, unit tests, and continuous integration.
Python · MIT · alpha-stage research software View on GitHub →
prCARS: phase retrieval and Raman-like signal reconstruction
A modular pipeline for recovering Raman-like information from CARS/BCARS spectra using Kramers-Kronig and Maximum Entropy retrieval, with non-resonant-background correction, preprocessing, denoising, and an optional neural-network retrieval interface.
Python · MIT · alpha-stage research software View on GitHub →
CARSGuard: validation and quality control
A reconstruction that looks plausible is not necessarily physically valid. CARSGuard scores CARS/BCARS spectra across four interpretable dimensions — physical realism, Raman consistency, artifact risk, and confidence — and produces warnings, recommendations, and reference comparisons rather than a single opaque quality number.
Python · MIT · alpha-stage research software View on GitHub →
VAE reconstruction under sample scarcity
Labelled data is the practical bottleneck in biomedical imaging. This project asks whether VAE-reconstructed images improve downstream classification stability or sample efficiency when labelled data are scarce, comparing classifiers trained on original versus VAE-reconstructed chest X-rays across training set sizes, with bootstrap confidence intervals.
Preliminary results are included and clearly marked as exploratory.
Python · MIT · v0.1.0 View on GitHub →
MedicalImageClassifier
A deep learning pipeline for medical image classification, covering preprocessing, model training, and evaluation.