No-Code AI
Quality Platform
Designing the interface between factory workers and AI-powered robotics — from a live inspection UI to a no-code training platform that let customers own their own AI without an engineer in the room.
- Role
- Lead Product Designer
- Team
- CEO · CTO · ML Engineers · Web Engineers
- Timeline
- 2020 – 2022
- Company
- Elementary Robotics
Background
Problem
Factory quality inspection was expensive, unfriendly, and impossible to customize. Existing machine vision tools required specialized operators, took weeks to configure for new product lines, and had no mechanism for factory workers to improve the AI over time. The technology existed — the interface didn't.
Goal
Design an end-to-end QA inspection system that factory workers could actually operate — and that customers could configure themselves. Two distinct products: a live inspection interface for the factory floor, and a no-code training platform that let managers create new inspection routines without engineering support.
Non-goal
This was not a product for ML practitioners. Every configuration flow had to work for a factory manager with no technical background. Internal tooling for the ML team was explicitly out of scope — if a feature required a data scientist to explain, we either simplified it or cut it.
Products, routines, and training progress in one view
The Routine Setup screen let managers see every product line alongside all of its inspection routines and each routine’s training state. Progress bars showing “9 labels til train” gave clear, actionable status — not abstract ML metrics. Managers knew exactly what labeling work remained before a routine could go live on the factory floor.
Add ML tools to each inspection position
Once positions were captured, managers could attach specific ML tools — Anomaly Detection, Barcode Scan, Color Match — to each camera view. Each tool trains independently on images from that position, giving the AI a dedicated model per inspection zone. No engineering support required.
Keyboard shortcuts cut labeling time in half
The original labeling UI required mouse-only interaction. Labeling a training batch of 500 images took close to an hour. Adding keyboard shortcuts — [1] Pass, [2] Fail, [3] Unknown — alongside visual hotkey indicators on each button cut labeling time significantly. The before/after also added inline pass/fail counts per label class, so managers could see their data balance as they worked rather than discovering class imbalance after training.
AOI-first vs. Tool-first — which mental model fits?
The original approach grouped all ML tools under one Area of Interest, creating a many-to-one mapping that confused non-technical managers. We inverted the model — each captured position gets its own dedicated tool setup. The relationship between “what the robot sees” and “what it’s checking for” became immediately legible without explanation.
Market Context
The existing market was Expensive, Unfriendly, and Hard to customize. We had to beat all three.
Expensive
Enterprise machine vision systems required six-figure contracts, dedicated hardware, and ongoing vendor support. Small and mid-size factories couldn't get in the door.
Unfriendly
Existing interfaces were built for machine vision engineers, not factory workers. Operators needed days of training before they could run a basic inspection.
Hard to customize
Configuring a new product line meant weeks of vendor integration work. Every new SKU was a new engagement. Customers couldn't own their own inspection logic.
Target Users
Inspector
On the factory floor running live inspection. Needs to see AI Pass/Fail judgments at a glance and correct them without breaking production flow. No technical background — the interface is their only window into the AI.
Factory Manager
Supervising multiple inspection stations simultaneously. Wants high-level quality signals and the ability to configure inspection routines for new product lines — without involving an ML engineer.
ML Engineer (Internal)
The team we were designing to replace. Every hour saved by self-service configuration was an hour our ML team could spend on model improvement rather than customer onboarding.
Design Principles
Lift from repetitive labor
The AI counted, measured, and flagged defects on every unit. Inspectors focused on judgment calls only — which reduced physical strain and cognitive load in equal measure. The interface had to make it obvious when a correction was needed.
Show the reasoning, earn the correction
A Pass/Fail verdict without context produced useless corrections — inspectors would override the AI without understanding what it had seen. Surfacing the inspection zone and the defect type turned corrections into training data.
Non-technical by design
Every flow in the no-code training platform had to work for a factory manager with no machine learning background. If a step required a data scientist to explain, we redesigned the step.
Process
Feedback Model
Explicit prompting won. Passive logging produced far lower-quality feedback because inspectors weren't aware their corrections were being captured as training data. When we surfaced the connection — inspectors became active participants in improving the model rather than inadvertent ones. Correction rate increased 3x after the prompt was added.
Confidence Display
We removed the number. In usability testing on the factory floor, inspectors fixated on the percentage and second-guessed correct AI calls when confidence dropped below their (arbitrary) threshold. The three-state model matched their actual decision model: act on it, reject it, or ask a supervisor. Cycle time improved by 15% after we made the change.
No-Code Onboarding
Modular setup was the right call, despite being harder to build. Training an ML model takes hours — sometimes days for a new product line. A single wizard forced customers to block off a day and stay in the product. Modular phases let managers start a capture session in the morning and return to configure the model after lunch. Three of our first five customers cited this as the reason they completed setup.
Outcomes
efficiency gain across customer inspection workflows
reduction in internal deployment cost per customer
new enterprise customers onboarded through self-service
“The no-code platform didn’t just reduce deployment cost — it changed the sales motion entirely. Customers could self-evaluate with a trial before any engineering engagement. That was the unlock.”