Designing a data platform so research can finally browse its own results
Role:
- product-design
- user-research
2022–2024
Switzerland
Summary
From access reserved to the terminal, where 86% of published data went unviewed, to a graphical interface that links data by provenance, with 94% of data viewed. I asked three backend engineers to measure usage, defended a real platform to structure, annotate, link and share data, and proved it could work with a test on a 20-neuron circuit. More than 2,600 researchers are now active each month.
Ask me
- Company
Blue Brain Project, EPFL
- Year
2022 – 2024
- Role
Product Design Lead and frontend engineer
- Team
The NISE team, 3 backend engineers, and the computational neuroinformatics and HPC teams
- Star metric
Data viewed from 14% to 94%
- Problem
86% of published data was never viewed, because it could only be reached through the terminal and an SDK, while scientists mostly wanted to see the links between data.
- Decision
Refuse to reduce the product to a library search with API endpoints, back that refusal with five years of usage data and qualitative research, then prove feasibility with a test on a 20-neuron circuit.
- Result
Data viewed from 14% to 94%, with more than 2,600 active researchers a month.
01
Context & business friction
Published data that almost nobody looks at
At the Blue Brain Project, a first usage diagnosis showed that 86% of Fusion’s published data was never viewed. It could only be reached through an SDK or the terminal, while scientists wanted to see the links between data. Internally, some wanted to reduce the product to a search and API endpoints. The full platform ran into three objections: cost, a priority judged lower, and data quality.
02
The Key Trade-off
What I pushed through, negotiated or refused
Some wanted to reduce the product to a library search and API endpoints to copy into a terminal. I defended a real platform, to structure, annotate, link and share data. In two to three meetings, I backed my position with interviews with scientists and five years of API requests, which showed the terminal was rarely used. I also refused to launch simulations from the web app: that called for a full application of its own, with its parameters and data validation before each simulation.
Group test sessions and one-on-one interviews with scientists, the evidence behind my position.
03
The Transformation
From project lists with no landmarks to exploration by data type and by provenance
Before, Fusion lined up two long lists of projects and studies, where the same entries repeated identically, and a menu of scattered tools: Search, Studio, Workflow, Admin, Kibana…
Now, search works by data type, morphologies, electrophysiologies, simulations…, each with its number of datasets. Useful items can be pinned.
Provenance finally links the data together. A Relations tab shows an entity, here a thalamic neuron, with its parent circuit above it, its siblings beside it and its 30 ion channels below, and a level slider climbs to the circuits that contain it. Next to the circuit preview, the “Built from” and “Used in” blocks list the linked datasets, upstream and downstream.
04
Business Impact & Proof
The measurable value generated for research
Fusion is finally being used: the share of data viewed rose from 14% to 94%, with more than 2,600 active researchers a month on the web app. The platform holds more than 12,000 reports, 20,000 experimental data items and 150 published studios.
And the proposal, contested at first, became a proof of concept approved by four team leads and validated on real data.
05
Critical hindsight & retrospective
What I learned and how mature I am
I would have added a data curation console, used by the BBP curation team, to show the life cycle of a piece of data rather than a frozen image. Users wanted to know when data had been created or updated, but time and people were short, so I left it aside.
I learned that usage data settles things better than opinions: five years of API requests and qualitative research resolved a debate that a designer’s opinion would not have.
One point remains open: the feasibility of the relations module at the scale of millions of resources, and getting it into production.
