Data Fusion by Qarnot

In 2018, I have had the opportunity to do an internship at Qarnot Computing, a french start-up, a green tech company. Qarnot create distributed computing data centers. They produce computing heaters, that uses computational resources (CPU/GPU) instead of resistors for producing heat. One could also say that they use the energy usually lost by a data computing machine to heat houses.

They now create other computing machines than heaters, and all the machines are equipped with sensors (temperature, humidity, CO2…etc). Those sensors can be use for new services as they provide data that Qarnot could use to help the inhabitants (e.g. advising to open the window whenever the CO2 level is very high). The sensors that are embedded on the machines respect the privacy of the inhabitant (hence no microphone, but a sound sensor), and Qarnot leaders want to develop services that continue in that direction.

My role there was thus to develop the basis of OASIS, the framework that allow getting proper data to develop intelligent services.
The problematic was thus to transform raw data, containing outliers, not saved with the same time frequency, geographical position, with potentially missing information, …, into workable data. This process is called Data Fusion.
Unfortunately, I can’t provide here any algorithm that we used/developed, as this was confidential.

However, this internship allowed me to understand how difficult it is to get a proper data set to perform machine learning on. Real world data is indeed very noisy, particularly in robotics, and data fusion is an essential layer (otherwise: “garbage in, garbage out“).