Case study — MSAID · 2021
Proteomics data is large, complex and unforgiving. How do you build one interface where a specialist and a newcomer can both run an experiment without breaking it?
The overview answers where a researcher stands: what is uploaded, what is running, where they were last. Storage and compute budget sit permanently in the sidebar — on this platform they are part of the work, not an invoice detail.
Click a screen to enlarge
AI analysis for proteomics research
The MSAID Platform analyses mass spectrometry data with AI-driven algorithms, automating peptide and protein identification so researchers get to the biology faster.
The science was never the bottleneck. The tooling was: fragmented desktop software, file names as the only organising principle, and workflows where one wrong parameter costs hours of compute and a repeat run.
MSAID needed a platform that made this accessible without softening the scientific rigour — usable by a proteomics specialist and by a lab member who inherited the instrument, on the same data, with results that stay reproducible.
At a terabyte per dataset, an interface is not decoration. It decides whether an experiment is repeatable.
Where scale meets usability
Product design across the platform
I designed the product end to end: information architecture, the file and experiment model in the interface, the filter and tag system, the experiment wizard, upload management and the visualisation surfaces.
Most of the work was learning the domain well enough to compress it. What is a raw file, what makes an experiment valid, which parameters can be defaulted and which must never be guessed for someone — those answers came from the MSAID scientists and then had to survive being turned into a form.
The interface therefore leans on structure rather than simplification: nothing was hidden that a result depends on, but nothing was demanded before the moment it matters.
Three decisions that made the scale workable
01
Creating an experiment means naming it, tagging it so it can be found again, and choosing an algorithm and version — decisions that determine the result and the compute bill.
So it became a guided sequence with a visible table of contents: name, tags, algorithm, then the settings that algorithm actually needs. Version selection is explicit — latest or pinned — because reproducibility is the point of the whole platform.
02
With thousands of files, a search box is not an organising principle. The platform got a filter model built for combination: chains of conditions over tags, dates, related experiments and metadata, each visible as a chip, all resettable in one action.
Tags carry the semantics the file system cannot — sample, instrument, patient cohort, QC state. They are editable in bulk, which is what makes them survive contact with a real lab where naming conventions change mid-project.
03
Datasets run from several gigabytes to a terabyte, so the upload manager was a design problem in its own right: aggregate progress, per-file state, remaining capacity, and a transfer that can be left alone without a person babysitting a browser tab.
At the other end, results had to be readable. The visualisation dashboard puts charge state distribution, identifications, precursor plots and volcano plots side by side, each bound to its experiment and exportable — so a finding can leave the platform without losing where it came from.
Power, without the ceremony
The platform turned complex proteomics workflows into something fast and reliable to operate: large datasets explored through filters and tags, experiments created without guessing, results visualised in one place — with the scientific rigour intact.
Nothing a result depends on was hidden; it was sequenced, so depth stopped being a wall.
Bulk-editable semantics were what let a real lab reorganise thousands of files after the fact.
Aggregate and per-file progress made a terabyte upload something a person can walk away from.
Keeping every visualisation bound to its experiment and parameters is what makes sharing safe.
In a scientific product, clarity is not the opposite of rigour — it is how rigour survives being used.
MSAID is where I learned to design for reproducibility: the interface's job is to make the conditions of a result visible and repeatable, not to make the result look effortless.
That is the same discipline I now apply to AI products. An algorithm choice, a version, a parameter set — or a model, a scope, a provider — are the things a person has to be able to check afterwards.