I build the ML systems behind precision oncology at 4baseCare, where I joined as a founding member of the data science team and built out its ML infrastructure from zero. My work spans three areas:
- Computational pathology — MIL and foundation-model pipelines (H-Optimus, Virchow2, UNI, CONCH) for biomarker prediction directly from whole-slide images, covering PD-L1, EGFR, MSI, and ALK fusion
- Liquid biopsy — cfDNA methylation panel design for minimal residual disease detection in colorectal cancer
- ML infrastructure — production pipelines on GCP (GKE, Cloud Run, Kafka, Prefect, Vertex AI) serving models at clinical scale
|
Pathology foundation models Domain-adaptive continued pretraining of a ViT-L/14 encoder on thousands of whole-slide images, incorporating efficiency techniques from Kaiko's Midnight architecture. |
Biomarker prediction from H&E Multi-model MIL pipelines (ABMIL, CLAM-SB, TransMIL) for predicting molecular biomarkers straight from tissue slides, benchmarked across multiple foundation model embeddings. |
|
cfDNA methylation MRD panel A 9-marker methylation panel for colorectal cancer MRD detection, validated in-silico against TCGA and positioned against existing liquid biopsy assays. |
cue An open-source terminal AI tool, built and maintained independently outside of work. |
ML / Data
Infra
Other
