COGNITE · 2026 · DESIGN TASK
commits
prototype builds
Overview
Engineers struggled to spot earlier abnormalities in their own data, and to trust that automatic detection would not miss new faults. I designed and built a working prototype where the AI detects and the engineer judges.
Role
Timeline
Responsibilities
Contents
01
Task: redesigning how humans & machines detect faults together
Machine learning algorithms can be better suited at monitoring & detecting abnormalities, while humans are (currently) better at recognizing which abnormalities are ok and which are not. Please watch this demo video and treat this as the “before” (manual) solution.
How would you go about redesigning this system for the LLM era? Please ideate and prototype a solution using AI prototyping tools.
Make reasonable assumptions where required. You can choose to focus on details of one part of the system, or a overall concept for the whole human-machine collaboration
OK behavior: lower turbine speed (black line) while low wind conditions (orange line) indicate normally functioning wind turbine
Unwanted behavior: sudden low turbine speed (black line) while steady wind conditions (orange line) indicate problem in wind turbine
Before: label a range by hand, train a model, then watch a status pill.
02
Typical user: engineer with significant technical knowledge (computer systems and industrial expertise) and abstract/analytical skills. Works in office environment, primarily at his desk but also significant meeting activity.
Imagine that the first round of informal user testing of this solution revealed that users struggled with 2 things:
Identifying earlier abnormalities in their own data
Trusting that the automatic fault detection would not miss any new faults
03
The brief asked for reasonable assumptions. These three shaped every screen.
Noisy desk. Eirik already gets alerts from many channels and people, so the tool adds as little as possible.
Desktop only. He works at a desk on a laptop or monitor, 1280px and up.
Three verdicts. Fault, Investigate or Not an issue. No warning level: the detector only finds sharp events.
04
The AI detects, Eirik judges.
Rules, not retraining. Every correction becomes a rule he can read and undo.
Show what was checked. Cleared dips stay grey, so trust comes from coverage, not silence.
Quiet by default. Browsing never starts a chat, and alerts go only where he chooses.
05
Problem 1. Users struggled with identifying earlier abnormalities in their own data.
Answer. The AI flags a mechanical fault before SCADA alarms and explains it from the data. Eirik confirms it and shares the section.
06
Problem 2. Users struggled with trusting that the automatic fault detection would not miss any new faults.
Answer. When the AI misses a drop, Eirik labels it. The AI owns the miss, finds another like it, and takes his correction.
07
I built it with Claude Code in small, checked steps. Each card is a step on the way.
01
Paper sketch first
The greeting, the graph, its timeline and an AI panel beside it.
02
Rules before prompts
Every decision lives in CLAUDE.md, read before each change.
03
The brief, read literally
Version 1: an edge-to-edge graph with the AI in a column.
04
Two blocks
Version 2: the sketch’s layout, graph and AI side by side.
05
Researching for references
Adobe Premiere’s timeline shaped the 30-day strip under the graph.
06
Research changed it
Curtailment, a planned slowdown, became the false flag to learn.
07
The Cognite brand
Three verdicts, quiet alerts, light and dark.
ITERATIONS
Version 0 · First build
The brief, read literally: a graph and a chat.
Version 1 · Full screen
An edge-to-edge graph with the AI in a column.
Version 2 · Two blocks
The sketch’s layout: graph and AI side by side.
Version 3 · Teaching loop
Every correction becomes something the AI learns.
Version 4 · Brand
Manrope, the Blue Fade, colour only for meaning.
Final
Three verdicts, quiet alerts, a mechanical fault explained.
08
PROBLEMS, ANSWERED IN A WORKING PROTOTYPE
Real. The interaction, the rules Eirik teaches and the alerts.
Simulated. The turbine data and the AI’s replies.
Next.
Review the UX copy of each flow.
Test the training flow with customers.
See how engineers use the AI, and whether they need it.
Validate notifications, to learn what should be on by default.
Find what to remove. It still feels overwhelming to me.
