← All work
Turbit AI monitoring, case detail with AI diagnosis
Case Study · Freelance · AI Monitoring · Turbit

Turbit, AI Monitoring
for Wind Energy

Turbit's AI watches thousands of wind turbines and detects anomalies before they become failures. My job: turn that analysis into an interface operators can actually act on, and rebuild the design foundation of the product along the way. Solo senior designer, end-to-end, from audit to concept to a design system built in one month.

Lead Product Designer Solo · End-to-End AI Monitoring Design System B2B SaaS
AI / ML Product Design Data-heavy UX Complex Workflows Design System Product Relaunch Figma Prototyping in Code

The problem: the AI is smart, the interface wasn't

Wind farms produce enormous sensor streams. Turbit's machine learning compares every turbine against a learned normal state and flags deviations early, days or weeks before conventional threshold alarms would fire. That's the product's superpower.

But the interface presenting those findings had grown engineering-first: dense tables, raw plots, and terminology that made every case feel like homework. Monitoring teams handle dozens of open cases at once. When each case takes effort to even read, triage slows down, and the value of early detection quietly evaporates. The design problem wasn't the AI, it was the distance between AI output and human decision.

Zoom: the case card, before and after

The case card is the atomic unit of the whole workflow. Operators scan dozens of them daily to decide: investigate, escalate, or dismiss. The original card front-loaded raw analysis, power curves, sensor plots, model output, and left the reader to work out what mattered.

Case card before redesign, dense technical detail
Before, every case opens as raw analysis. Status, severity, and next action are buried in the data.
Case card after redesign, decision-ready summary
After (concept), one glance answers: what happened, how confident is the model, what should happen next. The analysis is still there, one level deeper, for those who need it.

The principle: headline first, evidence second. The AI's finding is summarised in operational language, with severity and confidence visible at card level. Progressive disclosure keeps the full technical depth available without making it the entry price.

The relaunch concept: from fleet to case in three levels

Monitoring works at three altitudes: the whole fleet, a single park, a single case. The concept gives each level one clear job and a consistent visual language, status is always readable at a glance, and every level answers "where do I need to look next?"

Turbine fleet overview concept
Fleet level (concept), every turbine's state at a glance: healthy, watch, act. Anomalies surface themselves instead of hiding in tables.
Case list concept
Case queue (concept), triage view with severity, status, and ownership. Built for scanning dozens of cases, not reading one.
Case detail concept with AI diagnosis
Case detail (concept), the model's diagnosis, the supporting evidence, and the action path in one coherent view.
Park level view concept
Park level (concept), one wind farm, its turbines, and open cases in context.

Designing the AI workflows, not just the screens

A monitoring product lives or dies by its workflows around the model: how findings enter the system, how they're enriched, and how humans stay in the loop. Part of my work was mapping and designing these flows, including automation like extracting structured case data from unstructured inputs, so operators start from a pre-filled case instead of a blank form.

Automatic extraction workflow
Automated extraction, concept, unstructured input becomes a structured, pre-filled case. The human reviews and confirms instead of transcribing.
Whiteboard sketches mapping the park, windpark and turbine creation flow
Working the defect-tracking and setup flows on the whiteboard, mapping park, windpark and turbine creation before a single screen was drawn.

A design system in one month

The relaunch needed a foundation, and there was no time for a six-month system project. My approach: extract the system from the product instead of inventing it beside the product. I audited the existing interface, identified the real recurring patterns, and consolidated them into tokens and components, then designed the new concepts exclusively from that system, which stress-tested it immediately.

  • Audit of the live product → pattern inventory of what actually recurs
  • Tokens for colour, type, spacing, and status semantics, status being the heart of a monitoring product
  • Component library scoped to real needs: cards, data displays, status indicators, filters
  • Validated by building the relaunch concepts entirely from the new system
Turbit design library with tokens, buttons, inputs and status cards
The design library, concept, one source for tokens, buttons, inputs and the status cards that carry the whole monitoring product. Built to be themeable and code-mapped.

What I learned

  • In AI products, the model's confidence must be a first-class interface element, hiding it breeds either blind trust or blanket distrust
  • "Headline first, evidence second" scales: the same principle worked from card level to case detail
  • Extracting a design system from an existing product is faster and more honest than designing one in isolation, the product tells you what it needs
  • Solo freelance work on a technical product only works with engineers as co-designers: prototyping in code kept the loop short

Want the full story?

The shipped product is under NDA, I'm happy to walk through it in a call. Open to Lead, Principal & senior IC roles, Amsterdam or remote EU.