The hard problems we’re solving at Instrumental

Sam Weiss

I left my job building cutting-edge hardware at Apple over five years ago to start Instrumental because I love sinking my teeth into challenges where there are elegant solutions that may not be immediately obvious. Over the last five years, we’ve built a world-class team of engineers and technologists who have been tackling fascinating puzzles in Machine Learning, distributed system scalability, and UX with fundamental research and invention. And our solutions have real-world, tangible impact.

At Instrumental, we don’t make technology for technology’s sake: we make technology to solve real customer problems. If you’re considering a role at Instrumental, I’d love to share a bit of the core technological advancements we’ve made so far, and some of the challenges we’re tackling next. If these challenges are interesting to you, we’d love to hear from you: https://instrumental.com/careers.

Successful Production Implementation of Unsupervised Machine Learning

Machine learning is at the core of Instrumental products, powering real-time monitoring, on-demand searches, and discovery of novel anomalies in photos and other data. To be successful we must not only refine our techniques on fixed use cases: we develop our research and product direction together to craft an ML-assisted experience that makes our customers’ lives easier.

As an example, repeatably finding novel anomalies across many different customer product images has required significant research. We have now successfully productized a generalized anomaly/novelty detection algorithm, built on PyTorch, that does not require per-product training and that works with a cold start of just 30 total example images. This past year, we took this one step further by not just highlighting anomalies among images of the same type, but also automatically ranking these defect modes in order to surface the most relevant issues to our customers. These machine learning advancements power our recently released Discover AI product. 

Our users want to discover new issues quickly, with our best results at the top of the list. With Discover AI, our unsupervised machine learning algorithms identify both the most anomalous regions across our customers’ images and the most anomalous units within each region.

Running Production-Grade, Real-Time Models in An Extreme Few-Shot Setting

Eventually, after discovering a defect and finding its root cause, our users would like our in-factory edge devices to flag defective units in real-time.

Our main machine-learning challenge in the manufacturing setting is that we typically have very few examples of a defect to learn from when training a new model – this is called “few shot learning”. This challenge is particularly relevant as a customer transitions from smaller development builds to mass production. Despite having produced just a few thousand units during development, often with fewer than 10 instances of any particular defect type, our customers require confidence that our algorithms will behave well on their production line when millions of units are made!

Our first iterations of our algorithm required about 25 defect examples as part of a larger population to achieve our customers’ definition of “production grade performance” – for context, the best alternative on the market required 100s of defects. In 2020, we improved our models. Now, in most cases, just four defect examples are required to achieve production grade performance, and we’re not done yet!

Conventional computer vision algorithms require too many examples to be useful beyond the one or two highest prevalence defects – by being strong in the few-shot environment, Instrumental technology can be used deep into long tail of issues, as many as 100 defect types deep for some of our customers. That adds up to a lot of value!

Distributed Global Fleet of Highly Reliable Services

Instrumental is a critical part of every line we are on: our customers rely on our service to be highly available, repeatable, and robust to regular interruptions to their networks. A lot of our customers manufacture in China, which means our edge devices operate in some of the most challenging network environments in the world and with strict security requirements. At the same time, these edge devices collect and upload photos, transferring more unit data from our customers’ factories to our cloud platform than anyone has before. This is possible thanks to our remotely-managed, highly reliable software that operates even if the network is disrupted and seamlessly reconnects when the network returns. As a company, we have to excel at delivering the experience our customers expect of SaaS, with IoT-style devices, in an enterprise manufacturing setting.

assembly line

Our fleet of remotely-managed edge devices are workhorses on the factory floor – running machine learning algorithms at the edge to intercept defects when they happen.

Delivering the Magic with User-Centric Workflows

Instrumental’s application team talks a lot about “magic moments” – the instants when users get undeniable insights that they could not have gotten without us, frequently marked by an audible “wow.” This magic is often the result of carefully working machine learning into users’ workflows without them even realizing it. Crafting this UX is important because Instrumental’s users are mechanical engineers who are on a mission to ship great products; they need an easy tool to further their mission. Our modern decoupled web app enables users to create machine learning algorithms and gain insights from their predictions with a no-code interface that can be learned in minutes.

Our UX allows the powerful technology running under the hood to fade into the background, putting our mechanical engineer user’s mission first: are these bent springs causing poor antenna performance?

We’ve loved solving these challenges and seeing our customers not only build better products, but do so in less time with less waste. Instrumental’s value exists at the rare intersection of what is good for both our customers’ businesses and for a more sustainable future. There is still a lot of work to do and new problems to solve. If you’re excited about building technology that impacts the physical world – the products you and your loved ones use everyday – we’d love to hear from you.

Related Topics

Joel Garcia

Manufacturing Intelligence

Don’t miss a post, subscribe today.