rarelens — system architecture

One monorepo: scientific pipeline, API, UI, model serving and two deployment tracks. Arrows are data flow.

Clinical scientist

Enters a proband: a VCF plus the patient's HPO phenotype terms.

SvelteKit UI

Svelte 5 runes, adapter-node. Same-origin /api proxy, so no CORS.

web/

→

FastAPI

Cases, jobs, candidate ranking, decisions, case report. Pydantic v2 + SQLAlchemy 2.0 async.

api/app/ — 1,380 lines

Triage service

Weighted mean of four auditable components; a component with no evidence abstains.

app/services/triage.py

→

PostgreSQL 16

Cases, phenotypes, jobs, variants, predictions, decisions.

Alembic migrations

HPO reference

876,585 gene–phenotype annotations after ontology propagation; 11,829 terms with information content.

→

MLflow registry

LightGBM pyfunc carrying its own feature code, resolved by alias @production. MODEL_URI loads an artifact with no server running.

ml/

Nextflow DSL2 pipeline

bcftools normalise → Ensembl VEP → loader. Executors: local, Google Batch, Argo Workflows.

pipeline/

Deployment track A — serverless (default, ~£1/month idle)

Cloud Run

api + web, min-instances 0

→

Cloud Run job

Nextflow driver, started per run

→

Google Batch

VEP tasks on Spot VMs

Deployment track B — Kubernetes (behind a Terraform flag)

Pub/Sub

vcf-uploaded topic

→

Argo Events

sensor → workflow trigger

→

Argo Workflows

on GKE Autopilot

Both tracks run the identical pipeline code; the executor is a Nextflow profile, and ArgoCD reconciles the Kubernetes track from the same repository.