The architecture beneath the pipeline
Three engineering decisions run through everything we build: patient data never leaves the hospital, every recommendation must be explainable, and the heavy compute belongs on a GPU.
On-premise by construction, not by policy
Under EU GDPR Article 9, genetic data is a special category — processing it is tightly restricted, and hospital directors carry the liability for any breach. Our answer is architectural: the ingestion and reasoning services run as containers on the hospital's own network. No patient sequence is ever sent to a cloud API. Only de-identified, abstracted queries leave — to a public trial registry or a literature index.
Every recommendation is also framed as informational decision support. A clinician is in the loop at every step; the system answers questions and cites sources, it does not issue directives.
Accelerated AI Inference & RAG
NVIDIA BioNeMo Integration
Leveraging NVIDIA BioNeMo to predict 3D protein structures (AlphaFold2/ESMFold) and design highly affine neoantigens for personalized mRNA vaccines.
Genomic Foundation Models
Utilizing models like Evo 2, accelerated by GPUs, to score variants of unknown significance (VUS) and perform deep sequence-level reasoning.
Multi-agent Semantic RAG
A LangGraph architecture using local vector stores (e.g., Qdrant) and MedCPT embeddings to match patient genotypes with real-time clinical trials and literature without hallucinations.
Nextflow Bio-Pipelines
Highly parallelized, containerized execution of genomics workflows (WiGiTS, SAGE, LILAC) capable of running on local HPC clusters or bursting to cloud GPUs.
Every GPU workload, justified
We don't reach for a GPU by default — each workload below is GPU-bound because the alternative is either too slow to be clinically useful or physically impossible on CPU. Where CPU is the right tool, we use CPU.
Deterministic calling (SAGE/PURPLE/LILAC) already runs on CPU — but scoring the variants no database has classified needs a foundation model in the loop, and that doesn't fit a real-time CPU budget.
Structure prediction and docking (AlphaFold2/ESMFold-class models) to design and rank personalized vaccine candidates.
Low-latency, quantized on-premise serving — the copilot answers in seconds without a single call to a cloud API.
Thousands of interacting agents updated every frame — a CPU-bound simulator cannot run this at a clinically useful scale. Compiles to native CUDA on NVIDIA hardware.
Roadmap — evaluating GPU-accelerated retrieval and graph-traversal engines to scale the knowledge graph further.
A compute-first simulation engine, built to run anywhere
Our Agent-Based Modeling engine is written in C++20 with SYCL — a vendor-neutral compute standard that runs across GPU architectures. On NVIDIA hardware, it compiles to native CUDA for maximum throughput. By offloading massive spatial hashing and cell-to-cell collision detection entirely to GPU compute shaders, we eliminate CPU bottlenecks typical of legacy simulators — letting us model the tumor microenvironment at a clinically useful scale.
Compute / AI-driven solutions
Every simulation runs as GPU-accelerated compute — built for scale and speed, so the science isn't limited by hardware.
Built for scientists, not engineers
A visual editor lets researchers design and run simulations directly — no GPU programming or coding required.
Portable core, CUDA-tuned
SYCL keeps the codebase portable in principle — but every build defaults to NVIDIA's CUDA backend, where our own development, BioNeMo integration, and Evo 2 inference already live.
Built with