Every year, thousands of oncology patients endure chemotherapy regimens calibrated through trial and error. Miscalibrated doses cause severe toxicity or under-treatment. High-resolution genomic sequencing and MRI scans already exist in abundance, yet they sit in separate silos, never unified into a predictive model. The field needed an infrastructure capable of bridging molecular biology, anatomy, and pharmacokinetics in a single real-time pipeline.
The real obstacle wasn't raw compute. It was data coherence: making 150 GB whole-genome sequencing files, 3D vascular meshes extracted from DICOM, and pharmacokinetic models speak to one another on a single computational graph without sacrificing scientific fidelity. Initial CFD simulation times exceeded 18 hours per patient, making clinical use a non-starter. Each rare genomic variant demanded its own validation pathway.
Zeklar designed a directed-graph Nextflow pipeline that orchestrates genomic ingestion, PyTorch3D vascular segmentation, and the CUDA CFD solver in parallel stages, cutting end-to-end time to under four hours on an A100 cluster. For vascular meshes we trained a 3D segmentation model on 2,400 radiologist-annotated scans, reaching a Dice score of 0.91. The pharmacokinetic engine maps patient-specific SNP and CNV variants onto a Starling-Onsager model calibrated per patient. NVIDIA Clara provided the DICOM-to-tensor abstraction that made KAIROS compatible with existing hospital PACS.
Discovery & data modelling
We mapped clinical data flows, defined the unified genomic-physiological schema, and locked validation criteria with the research team.
Segmentation pipeline
We trained the 3D MRI segmentation model, integrated NVIDIA Clara, and automated vascular mesh production end to end.
CFD & pharmacokinetics
We built the CUDA solver, wired Starling-Onsager to per-patient genomic variants, and optimised for multi-GPU throughput.
Validation & deployment
We ran the 340-patient retrospective study and deployed on a hybrid on-premise/cloud infrastructure with native PACS integration.
KAIROS aims to cut chemotherapy adverse events by ~30%, to be demonstrated through clinical validation with partner oncology centres. Today it is an in-silico research prototype, eligible for digital-health funding programmes; it is not a medical device.