Projects

Things I have done at work, at school, and beyond!

AI-based Cell Instance Segmentation for Quantitative Phase Imaging

Languages: Python • C++ • Batch • Bash

Frameworks: Tensorflow • PyTorch • Detectron2 • TensorRT • ONNX • Vertex AI • Lightning AI • OpenCV • WSL • Qt

Domains: AI • Computer Vision • Cell Biology • Bioprocess Engineering • UX

Destription: Through local and cloud-based workflows, I trained Detectron2 instance segmentation models to distinguish between cellular states in images taken from Phi Optics' Quantitative Phase Imaging microscopes. To maximize performance, I implemented convolutional backbones in TensorFlow and Pytorch and created robust dataset pre-processing workflows. In Python, C++ and Qt, I also developed a test application integrating ML inferrence with a GUI using ONNX and TensorRT engines. This app has been instrumental in assessing interest from investors and biomanufacturing firms for Phi Optics' novel ML-based bioprocess monitoring & control software.

AI-based Phase Separation Phenotype Segmentation in Neurons

Languages: Python

Frameworks: Tensorflow • HPC • Slurm • OpenCV

Domains: AI • Computer Vision • Neurobiology • Aging • Biochemistry • Cell Biology • Phase Separation

Destription: To classify and analyze visually similar but chemically distinct images of sub-cellular structures in cultured neurons, I used Python, Tensorflow, and Slurm-based HPC systems to train UNet models with InceptionV3 backbones. With minimal training input (6 manually annotated images), my models achieved 96% pixel-level classification accuracy and over 85% semantic segmentation IoU, significantly accelerating our understanding of age-related changes in neuronal metabolism and innovating new ways to analyze phase separation in cell images.

ProDiVis: Protein Ditribution Visualization

Languages: Python

Frameworks: Jupyter • OpenCV • Matplotlib

Domains: Neurobiology • Oncology • Image Analysis • Biochemistry • Cell Biology

Destription: To investigate how differentially modified proteins could cause a pro-cancer metabolic switch in cellular models of Glioblastoma Multiforme, I co-developed a confocal microscopy-based z-stack normalization and visualization pipeline dubbed Protein Distribution Visualzation (ProDiVis). The pipeline leverages common python packages like Jupyter, OpenCV, and Matplotlib to significantly enhance biological insights at deep focal planes in 3D imaging.

Cloud-Based DNA Sequence Quality Control for Personalized CAR-T Therapy

Languages: Python • SQL • R

Frameworks: AWS • GCP • Jupyter • Pandas • RStudio • VCF

Domains: Bioinformatics • Oncology • Cloud Computing • Statistics • Visualization

Destription: Implemented genome fingerprinting and comparison into a custom tissue processing pipeline for patient-personalized CAR-T therapy in the nation's 11th largest health system. Integrated into existing AWS workflow as a quality measure to conclusively determine whether a tumor sample comes from a patient or not.

  1. Created and fingerprints from the hundreds of tumor & normal tissue samples.
  2. Assessed quality control efficacy with statistical and visualization methods.
  3. Enabled integration into AWS pipeline for high-throughput quality control.

Back to top.