Case studyEdge AI
Making sensor intelligence easier to see.
A visual language for turning complex sensor data and machine-learning models into tools engineers can reason about.
- Client
- Reality AI, part of Renesas
- Industry
- Edge AI and industrial sensing
- Product
- Reality AI Tools
- What we did
- Developer experience, data visualisation
About the client
Reality AI builds software that turns raw sensor data, such as vibration, sound, current, and radio signals, into machine-learning models small enough to run on low-power microcontrollers. Its flagship product, Reality AI Tools, supports engineers from collecting data to deploying a model on a device.
Renesas Electronics acquired Reality AI in 2022 to bring this kind of sensor intelligence to its microcontrollers and processors.
Source: Renesas, June 2022Overview
Embedded machine learning asks engineers to reason about three hard things at once: raw signals they cannot see, models whose behaviour is opaque, and devices with very little memory or power. Our work focused on making that loop visible.
Business requirement
Give engineers a clear line of sight from data to a deployable model: inspect signals, understand what drives a model’s decisions, and weigh accuracy against memory and power before anything goes onto a device.
Solution
We focused the product experience on the engineering loop. Signal views let engineers look at recordings, compare classes, and spot noise or mislabelled data before training. Model views show which features a model relies on, so its decisions can be checked against what an engineer knows about the machine.
Deployment trade-offs are laid out side by side, with accuracy, memory, and compute cost for each candidate model, so choosing what to ship becomes a visible decision rather than a guess.
HighlightsWhat the product gets right
What the experience is built around
See the signal
Views that make raw sensor recordings readable, so problems in the data are caught before training.
Explain the model
Feature views that show what a model relies on, so engineers can trust it or challenge it.
Trade-offs in the open
Accuracy, memory, and compute for each candidate model, side by side, before choosing what to deploy.
Built for engineers
Dense, precise interfaces that respect expert users instead of hiding detail behind simplicity.
ChallengesWhat made it hard
Challenges we worked through
- 01
Data people cannot see
Vibration and sound are invisible. Turning them into views an engineer can read quickly was the heart of the work.
- 02
Opaque models
Engineers will not ship what they cannot explain. The product had to show why a model decides what it does.
- 03
Hard device limits
A model that is accurate but too large is useless on a microcontroller. Constraints had to be visible from the start.
- 04
Expert users
The audience knows signal processing deeply. The interface had to be precise and information-dense without becoming confusing.
ApproachHow the work ran
Our approach
We worked around the engineering loop itself, treating each step from data to deployment as a place where the right view could save hours.
- Step 1
Follow the loop
Map how an engineer goes from a recording to a deployed model, and where they get stuck.
- Step 2
Design the views
Create visual languages for signals, model features, and deployment cost.
- Step 3
Prototype with real data
Test the views against real sensor recordings, where noise and edge cases show up.
- Step 4
Refine with experts
Tighten density, labels, and interactions until expert users can move fast.
Where it landed
A developer experience that gives technical teams a more direct way to explore, validate, and put edge intelligence into production.
A visual language for sensor data and models that can carry across the wider set of tools.
