Case study Genomics · Clinical and research SaaS · Designed 2021 to 2022 · Written 2026

Designing for petabytes: clinical-grade interfaces on top of genomic data.

Seven Bridges (now Velsera) is a cloud workspace where biomedical teams store, analyse and jointly interpret genomic data. This is a read of the product architecture I worked inside: how cohort discovery, reproducible pipelines and lab intelligence get turned into surfaces a wet-lab researcher can actually use.

RoleSenior Product Designer, Seven Bridges / Velsera
DomainBioinformatics · Clinical diagnostics
UsersResearchers, bioinformaticians, lab staff, compliance
MethodEmbedded product design + systems analysis
1M+Participants addressable in a single cohort query
21,600Phenotypic fields in the searchable data dictionary
550+Tools and workflows exposed as a public app library
2 to 5sCohort definition at interactive speed

No product screenshots. The platform is under NDA and handles regulated patient data, so nothing in this study shows the real interface. Every visual below is a diagram I drew for this write-up to explain the architecture, the flows and the numbers. Figures quoting scale or timing come from Seven Bridges' published benchmarks.

In 30 seconds
  • The problem: genomic data is too large to move, too complex to browse, and too regulated to guess at. Three constraints, one interface.
  • The design work: a shared analysis loop, an interactive-speed cohort builder, and auto-generated GUIs that let developers ship interfaces without designers in the loop.
  • Worth stealing: reproducibility rendered as interface, honest performance, and progressive disclosure from Run button down to API.
Four products, one substrate Platform · ARIA · GRAF · RHEO. Same projects, same files, same permissions underneath. Platform Projects, data, 550+ CWL tools in the browser ARIA Genomic + phenotypic cohorts at 1M+ scale GRAF Graph reference genome, secondary analysis RHEO Python ADK, GUIs generated from code Co-located compute and data Analysis runs next to the petabytes. The worst step in genomics UX, moving the data, is removed rather than optimised. Security, access control and audit Identity, managed ingress, provenance and audit trails: a structural layer, drawn in the architecture, not a badge in the footer. DIAGRAM DRAWN FOR THIS CASE STUDY · NO PRODUCT IMAGERY UNDER NDA
FIG. 01 The product map. Four named products sit on one substrate, which is why a cohort defined in ARIA can be analysed on the Platform without an export dance.
01 · Context

Three constraints that decide every screen.

Genomics UX is shaped by physics before it is shaped by taste. A single whole-genome sample is hundreds of gigabytes; a study is petabytes. That means the data cannot come to the user, so the analysis has to go to the data, and the interface has to make a remote, long-running, expensive computation feel like a document you opened.

The second constraint is comprehension. A phenotypic store with 21,600 fields is not browsable. Nobody scrolls to a cohort. The interface has to be a search-and-refine instrument, and it has to answer fast enough that defining a cohort feels like thinking rather than like submitting a job.

The third is trust. This work ends up in clinical diagnostics and regulatory submissions. If a researcher cannot see which tool version produced a result, the result is not usable, no matter how good the chart is.

02 · Users

Four people who need the same objects in different clothes.

The hard part is not that these personas want different features. It is that they want different renderings of the same underlying object, and a handoff between them is where projects die.

GUI ONLY CODE ONLY Wet-lab researcher Runs validated workflows. Wants the Run button, not the CWL underneath. Served by: public apps Bioinformatician Builds and versions tools, tunes cost and runtime. Lives in editor, CLI, API. Served by: revisions, app IDs Automation developer Wraps whole pipelines in Python ADK scripts. Ships to non-coders. Served by: generated GUIs Data steward The silent persona the architecture is built for. Access, audit, compliance. Served by: provenance One object model: project · file · app · task · cohort Every persona reads and writes the same records. The design job is the rendering, not a second product. DIAGRAM DRAWN FOR THIS CASE STUDY
FIG. 02 Personas placed on a GUI-to-code axis. The interesting design surface is the seam between neighbours, not the ends.
The core UX thesis

Move the analysis to the data, and make reproducibility a visible object on screen. Everything else is a rendering decision.

03 · The flows

Three loops that carry the whole product.

The platform looks enormous from the outside. In practice almost all value passes through three loops, and each one has a single moment where the design either holds or collapses.

Flow A · the analysis loop Projectscope + permissions Dataupload or public sets App550+ library or own CWL Taskinputs + instance size Runcompute at the data Inspect results, then iterate the revision preview in browser · metadata, raw and visual encodings side by side THE FRAGILE MOMENT: THE WAIT. A RUNNING TASK MUST STAY LEGIBLE, CANCELLABLE AND PRICED.
FIG. 03 The loop researchers live in. Revisions are first-class, so iterating never overwrites the run someone else cited.
Flow B · cohort discovery, with the clock on it Published benchmarks against a 600,000-patient phenotypic store and 140,000 whole genomes. Search 21,600 fields ~5 s Define 100k cohort ~2 s Statistical charts <1 s Export precise subset ~1 min Export broad subset ~15 min EXPLORATION IS INTERACTIVE · THE ONLY SLOW STEP IS THE ONE THE USER CHOSE TO COMMIT TO
FIG. 04 Where the seconds go. Spending the user's patience on export, not on exploration, is the single best decision in the cohort experience.
Flow C · the persona bridge Developer writes ADK script whole pipeline as Python, up to 80% less code Start page generated the GUI comes from the code, no interface build step Researcher presses Run one button, full provenance preserved end to end Results and failures return to the developer as the same objects, so the loop closes without a ticket. REMOVING A HANDOFF BEATS DESIGNING ONE
FIG. 05 Generating the interface from the pipeline definition is the highest-leverage move in the product: it deletes the step where developer output waits on design.
04 · Design decisions

What the UX work actually consisted of.

Reproducibility rendered as interface. Revision numbers, CWL version, contributors and app IDs live in the page header rather than in documentation. A scientist can answer "what produced this?" without leaving the screen, which is the precondition for the result being citable at all.

One saturated call to action per surface. These pages are dense with metadata, and density is fine, as long as there is exactly one thing that is obviously the next action. Run is the only strong colour on the page.

Progressive disclosure from button to API. The same objects are reachable through GUI, CLI, notebooks and API. A team can start in the browser and graduate to code without re-learning the model or migrating projects.

Lab intelligence as a dashboard, not a report. For diagnostics work the interface had to show sample flow, QC metrics and turnaround in near real time, so a lab manager can spot a stalled batch in the morning rather than in a weekly PDF.

Honest performance. Publishing the slow path next to the fast ones sets expectations before the spinner does. In scientific tooling, credibility is a feature, and overclaiming costs more than a fifteen-minute export.

05 · What held and what did not

An honest read, including the parts I would redo.

Held: co-location

Treating "the data never moves" as a UX decision rather than an infrastructure one removes the worst step in genomics work instead of decorating it.

Held: provenance

Versions and IDs as visible objects gave a regulated audience something to trust, and gave support a shared vocabulary with users.

Held: persona bridging

Generated GUIs and a public app library convert bioinformatician output into researcher-usable surfaces without a handoff.

Did not: naming

Platform, ARIA, GRAF and RHEO is an org chart projected onto users. Nothing told a new team which product to start with. A persona-by-task chooser was the obvious fix and never shipped.

Did not: showing the work

The best interface in the product, the reproducibility chrome, is the one thing competitors cannot copy, and it was almost invisible outside the app. Marketing surfaces led with abstract heroes instead.

Did not: onboarding depth

Power users were well served, first-week users much less so. The gap between "I have an account" and "I ran something meaningful" stayed wider than it should have.

06 · Takeaways

Four principles I carried out of this domain.

Physics first

In data-heavy products, the constraint chooses the interaction model. Design the interface after you have decided where the compute lives.

Trust is a component

Provenance, versioning and audit are UI work, not compliance paperwork. If a user cannot see why a number is true, they will rebuild it elsewhere.

Delete handoffs

The most valuable thing I saw here was an interface generated from code. Removing a step between two personas beats designing a smoother step.

Be slow on purpose

Publish where the time goes. Users forgive a long operation they chose; they do not forgive an exploratory click that stalls.

Confidentiality note: Seven Bridges and Velsera work is covered by NDA and involves regulated patient data, so this case study contains no product screenshots. All figures are original diagrams drawn to explain the architecture and flows. Quantitative claims come from Seven Bridges' publicly published platform benchmarks and product pages.

Next case study

Right-sizing PLM and MRP for hardware teams