Case Study · CSIRO · AI Product Design · TRL 2→7 · 2023-2026

Sciansa: AI for Scientists

Experimental biologists needed AI they could check, not just AI that gave answers. I led design of a multi-agent platform that shows its reasoning and lets scientists steer each step.

TRL range 2→7
Scientists reached 3,000+
Workflows run 12,000+
Sciansa AI platform interface

My role

Principal Product Designer. I led problem definition, concept development, testing, iteration and final design of the MVP.

The team: Product Managers x 3, Engineers x 12, Researchers x 2. I led the design team, where at different phases there were 2-3 people.

Result: 3,000+ scientists, 12,000+ workflows run, as of March 2026.

The problem

Experimental biologists want to identify key genes and understand biological pathways. The tools for this assume command-line skills and access to supercomputing.

Predicting gene function by hand is slow and often gives low-confidence results. Lab validation is expensive, so a bad prediction costs real time and money. Data also sits across disconnected tools.

The design question was about access and trust equally: how do you give scientists AI capability without asking them to take its results on faith?

What the research changed

We worked with 100+ domain experts through co-creation sessions and user research over a 2 year period. Scientists told us they needed three things:

1. To see the AI's reasoning.
2. To string workflow steps together themselves.
3. To get a clear explanation of how each result was generated.

Key quotes: “It has increased my Bioinformatics productivity tenfold”, “It directly aids in answering my research questions”

Sciansa design process stages

Design for surfacing agent orchestration steps.

Key decisions

1. Show the plan before anything runs

Sciansa drafts a workflow from a natural language request, and the scientist reviews each step before running it. We chose this to help researchers surface relevant agentic tools for their workflows. The cost was more steps before a first result, but the gain was that users quickly developed a mental model of the platform's offerings.

2. Make the evidence part of every step

Each step has its own context, output and notes, plus an explanation of what the agents were doing. The addition of the 'notes' feature enabled researchers to have finer control of the context they needed to support the next step in the workflow. Testing demonstrated that these types of context controls were critical for workflow usability.

3. Keep an honest record of what ran

Once a step has run, its prompt becomes read-only. This mattered as the agentic workflow steps became dependencies for downstream steps. This feature also enabled scientists to create reproducible workflows. The cost was less flexibility to edit agentic steps that had run.

Sciansa workflow with steps showing context, guardrails, reasoning, outputs and evals

Sciansa concept screen.

What happened

The early research held up beyond the prototype. By March 2026, 3,000+ scientists had run 12,000+ workflows on Sciansa. Since the MVP, bespoke agentic services such as Genome Assembler have changed how bioinformatics is done at CSIRO, and automating workflow pipelines is saving scientists real time.

What I'd do differently

If I was to design this platform again, I would lean more into IDE interaction patterns. My brief in 2023 was to create an AI platform for all types of scientists. For this reason, I didn't want it to feel like an 'engineering' tool. I did align some of the platform's features with patterns in IDEs, but with what I know now, I should have aligned it further.