Skip to content

Built like infrastructure, not like a script collection

Biomolecular NMR analysis is fragmented across specialist tools, each with its own formats, parameters and ordering constraints. Assembling a correct, reproducible pipeline demands expert knowledge and manual bookkeeping that is almost never captured. This is our answer to that.

Forty services, one contract

Every scientific capability is an independent HTTP service with a published OpenAPI specification — JSON in, JSON out, one engine per service. ChemEx, relax, ParaMagPy, pyDIFRATE and the rest sit behind that uniform contract, alongside the ten engines we wrote ourselves.

That uniformity is what makes everything else possible. Because each endpoint is already specified, the assistant's tool definitions are generated from those specifications rather than hand-written, so its capabilities cannot drift from the platform's.

Pipelines are executed by a durable workflow engine, not a shell script with a timeout. A catalogue of one hundred and eleven named workflows runs with checkpointing and heartbeats, so a six-hour model-free fit survives the things that kill long-running jobs.

Each pipeline family collapses its variants into one definition with multiple entry points, which is why you can enter at spectra, at a peak list, or at fitted intensities without a different pipeline for each.

Correct by construction

One toolchain, reachable two waysTwo entry points — a spectroscopist using web forms, and a newcomer using conversation with the assistant — both feed into a single workflow engine, which writes to a single result store. The same analysis, the same provenance, whichever route was taken.Web formsthe spectroscopist configuresConversationthe assistant does the legworkOne workflowengineOne resultstorewith provenance

Large language models can plan multi-step tool use, but “trust the model” is not acceptable for quantitative science. So scientific correctness does not depend on it.

A deterministic directed-acyclic-graph guardrail, derived from the platform's canonical workflow definitions, gates every operation on the prior existence of the artifacts it requires. Fitting before assignment transfer, or model-free before a correlation time, is structurally impossible. The model chooses what to do; the graph decides what is allowed.

Because the assistant and the web forms drive the same workflows and write to the same result stores, an expert configuring a pipeline by hand and a newcomer describing it in words reach an identical, provenance-tracked result — without us maintaining two implementations.

Provenance on every result

Derived parameters are frequently published without a machine-readable trail from raw data to result. Deposition standards address data exchange; they do not address analysis provenance.

How every result traces back to the raw dataA directed graph. A raw FID becomes a spectrum, which becomes a peak list. The peak list branches to a peak series and to an assignment. The series yields relaxation rates and exchange parameters; the assignment yields chemical shifts. Relaxation rates yield model-free order parameters, and chemical shifts yield backbone dihedral angles. Every arrow is a recorded step, so any result can be walked back to the FID it came from.FIDSpectrumPeak listPeak seriesAssignmentRelaxation ratesExchangeChemical shiftsModel-free S²Dihedrals φ/ψ

Every result is an envelope recording the identifiers it consumed and produced. Full lineage can be reconstructed back to source spectra and structures, which is what makes a published number auditable rather than merely reported. Results export as NEF, so they leave in a community standard format rather than ours.

Run it wherever your data has to live

One Helm chart

An umbrella chart with forty-four subcharts, including a bundled database, workflow engine and identity provider, so a complete stack installs with no managed services required. Kubernetes 1.24+ and Helm 3.8+.

Any Kubernetes

GKE, EKS, AKS, OpenShift or bare metal. A cloud-agnostic Terraform module covers the wiring, with worked examples for AWS, Azure and GCP. The platform differences reduce to a storage class name and an issuer URL.

Your identity provider

No user directory of our own. Every service verifies OIDC tokens against your issuer's JWKS — Entra ID, on-prem Active Directory, Okta, Keycloak, Cognito. No SDK, no vendor library, no directory sync.

Air-gapped installs

Every image can be mirrored into a private registry and rewritten at install time. Sites that cannot reach the public internet are a supported configuration, not an exception.

Portable storage

An object-store abstraction over cloud buckets or plain filesystems, so GCS, S3, Azure Blob or Files, NFS and ordinary volumes all work identically through Kubernetes CSI.

Authorisation you can audit

Identity comes from your provider's token and nothing else. What a person may touch comes from group membership, roles and dataset grants held in the database, with optional row-level security enforcing the same answer at the row.

Licences respected

An engine whose licence forbids redistribution — ModelFree — is packaged as a separate chart held back from distribution. A self-hosted install neither lists nor runs it; it is available in hosted deployments.

Your data stays yours

Nothing about a telemetry destination is compiled into an image. Deploy inside your own boundary and results never leave it.

Deploying inside your own infrastructure, or evaluating this for a facility? Talk to us — we will tell you honestly whether it fits before you spend a week finding out.