How to love machine learning with Unplugg

How to love machine learning with Unplugg

by Miguel Tavares -

Why unplug(g)?

Over at Whitesmith, we try to make the most of our work for it to be impactful on the real-world. This translates into delivering solutions to real-world problems like mitigating remote workers’ cabin fever or facilitating usually complex processes as bridging the gap between donors and social institutions.

This is a deliberate and conscious action, both driven by our company’s mission and each individual employee’s conviction that we all need to chip in to solve the world’s problems. Reflecting this motto is our oldest product: unplugg.

Shifting direction ever so slightly throughout the years (you can read more about that journey in another blog post), Unplugg’s purpose remained unwavering: making energy consumption “intelligent” by translating consumption metrics into informed suggestions, actionable information and patterns.

Why energy? Because since 2012 we believe it is where the most impact can be achieved by the tiniest progress, and therefore where there is the most for society to gain. We are not the only ones that think this away[^1], but as every player in this field, we need to figure out what benefits our particular skill set can bring to the domain.

This can be a tough endeavor, with a multiplicity of innovative energy sub-domains that require attention - renewable energy, smart grids, battery tech, CO2 emission reductions, just to name a few. But we’re a software company, after all, used to dealing with code, information, and data. Lots of data. And that’s good because we’re not alone in thinking that Machine Learning is the way to go[^2].

Unplugging MLaaS

One of the most defining traits we found when developing Unplugg was how little changed from distinct energy use-cases. Whether the client is an IoT startup or a massive utility provider, the needs are similar: forecasting, disaggregation, automation rules, these always kept popping up as major needs in every scenario. It doesn’t make sense that these need to be custom built for every player out there, there’s no need to keep re-inventing the wheel.

So we built Unplugg as a Machine Learning-as-a-Service (MLaaS) platform, making these common models equally available for all energy data analysis cases that could possibly need it. Not only that, but we kept the technical cost as low as possible: unlike other MLaaS platforms where the models are available but need to be trained individually, Unplugg provides fully trained models (or with no previous training necessary) ready to plug-and-play through a simple API to run ML tasks.

This is done both to reduce friction to adopt our models, and also to provide complex models to players that typically couldn’t access them - massive amounts of energy consumption data are not easy to come by if you are an IoT startup launching a new energy-related gadget.

After all, we’re not really looking at energy, we’re analysing human behaviour.

Moving beyond energy

But energy is only the beginning for Unplugg. Let’s take a step back and look at the bigger picture: the intelligence we are trying to bring to the energy domain is not that specific. Analysis of time series data is not exclusive to energy: sensorization, fitness, medical tests, temperature control, telecommunications, all these domains produce similar datasets that could benefit from similar intelligent models - forecast, pattern recognition, and action automation.

Before we continue, yes, we are aware that there are no free lunches - we can’t create a single general model and expect it to perform splendidly across very distinct problems.

However, looking closely at these domains, we see the problems to be solved are not that detached from the problems we attempt to solve at Unplugg. After all, we’re not really looking at energy, we’re analysing human behaviour. The metrics collected on energy consumed reflect human interaction with the environment, not energy itself. And those patterns follow their human creators, reflecting their habits, their routines, whether it is when they use most of their electric appliances or when they go for a run. The underlying producers of data - human routines - are the same.

This means that the models we’ve been using so far are not only useful to energy producers and consumers but to a vast array of human-data producing enterprises. As long as there is time-series data of human behaviour, it can be forecast, split into smaller, repeated patterns, and grouped and categorized with other similar behaving people, regardless of what the underlying actions represents - electricity, water, gas usage, mobile phone calls and data usage, temperature readings, fitness logs. All of these have the potential to let us know how people behave and, more importantly, how they are most likely to act in the future.

Unplugg your products

But in the end, what matters is: how can Unplugg make your products smarter? Maybe it can help deliver forecasts on user activity for your activity tracker. Or maybe make your smart meter help identify which appliances are turned on? The only way to know for sure is by trying it out.

Right now Unplugg’s forecasting API is open for testing (you just have to ask us for an API key here) and validation, and over the next few weeks we’ll be posting more detailed analysis on what our models can help you achieve in energy analysis, temperature control, and more domains.

In the meanwhile, keep yourself in the know by subscribing to Unplugg’s mailing list.

Think you have a killer idea that could benefit from similar machine intelligence? Let us know, we’ll love to spend some time discussing how to make the world a smarter place.


[2] </small>


Miguel Tavares

More posts

Share This Post

Right Talent, Right Time

Turnkey technical teams and solo specialists
when you need them for greatest impact.

Work with us