Outline:

• Usage of Data Management and Artificial Intelligence

• Energy Management System

• Goods Categorization System

• Search System for Microelectronic Components

• Miscellaneous Projects

There is a prevalent opinion that data is the new oil. At Sonerim, we believe more in what David McCandless said in his 2012 TED talk about data being the new soil: you need to take care of data to gain your AI-based crops. In this article, you will read about client success stories that demonstrate our approach to solving big data challenges in a modern AI-based world.

1. Energy Management System

Our first success story was a game-changing environmental project dealing with energy issues in Australia. One of the earliest brainchildren of Sonerim, this software was developed for managing solar panels at the request of our first significant client. This client manufactured a special device—a controller—that was installed in small households and big energy farms with solar panels, windmills, and other energy-generating equipment. 

As you might know, one of the biggest challenges in remote Australian regions is the frequent absence of power and its unstable supply; power failures can last for several days. The majority of households generate electrical energy for their needs and can even sell it. However, they don’t understand what volumes could be sold at a definite moment, and an unpredictable supply prevents them from earning more money. People have all the necessary equipment for power generation but can’t determine a timely action. They face a dilemma: either (a) they can sell power, or (b) they must charge their accumulator to the maximum capacity. That’s where we stepped into the game and helped solve the dilemma. 

We developed custom software for a controller and devised a special system for collecting data. It allowed for accumulating, monitoring, and managing power production, consumption, and redistribution in an efficient way. 

The energy management process included the following stages: 

1. First, a specialist visited a household and installed a controller. It received a large amount of data from solar panels and accumulators, information about the standard power consumption of the household, current prices in the network, and so on. It also received data from weather forecasts and notices on local websites and forums about impending power cutoffs. So, at this first stage, the controller collected large amounts of data about the power generation process (generated and consumed amounts, voltage charts, voltage surges, etc.) and some external conditions. 
2. At the next stage, all of that data was processed, and the software gave recommendations on what could be done with the generated power. The first option was to charge an accumulator to the maximum capacity; the second was to sell the generated power completely. If good weather was forecasted to last several days and no power cutoffs were in view, the household only had to charge its accumulator up to 70% of its capacity. 

The project challenges were the following:

• Really large amounts of data for processing. In a power generation process, a controller receives a great number of variables and sends them to a server every 30 or 60 seconds. The number of controllers in a medium-sized household could reach thousands. 
• As such, the controllers weren’t very smart. They collected data and sent them without checking for successful delivery. If some data were lost, it was final; the controller didn’t save data in its memory. 

While developing the software for this project, we assigned one controller to be an isolated, maximally stable data receiver. It was adjusted only once, without subsequent amendments, for receiving and storing all input data. 

What models and approaches we used:

1. We analyzed data, built graphs, and gave forecasts about power consumption and its further usage with the help of artificial intelligence models. 
2. For product marketing, we also used image recognition. We analyzed satellite pictures to determine houses with solar panels and delivered our product’s ads to their mailboxes. 

The technologies we employed during the project completion: 

• Amazon
• Docker
• Casandra, PostgreSQL
• React.js
• Java Android
• Java, Spring
• Kubernetes
• PHP
• Swift
  

2. Goods Categorization System

Sonerim developed this project for a big company from the Netherlands, a renowned manufacturer of household goods encompassing hundreds of thousands of items (drapes, tables, chairs, and so on). This company sells its numerous products on nearly all prominent online marketplaces. 

Having a huge assortment of goods, our client faced a major challenge—the extreme difficulty in maintaining and updating the actual status of product information. As you might know, every online marketplace has its specific structure, descriptions, numerous headings, and subheadings; a classification of, for instance, a single chair could meet with the following obstacles:

• One marketplace might have ‘Kitchen Goods→Furniture→Chair’ categories;
• Another one might have ‘Furniture→Garden→Tables & Chairs’ categories;
• A third marketplace might have one more category in addition to the above.

A physical product (in our example, a chair) is the same in every marketplace, but its categories can differ; sometimes, there are two different categories for one product within the same shop. That is why the task of adding a new product to the correct category in the assortment was challenging. The process meant manually auditing all categories in every marketplace and comparing them with the existing goods. This overcomplicated categorizing of new products required a full-fledged department of 40-50 specialists. 

What did we undertake to simplify our client’s life? We developed a system that:

1. Was integrated with our client’s enterprise resource planning system (data on manufacturing, supply chain, finances, procurement, etc.) to give an opportunity for monitoring their entire stock; 
2. Employed artificial intelligence to determine a category for every new product in online marketplaces and to verify and update the existing ones by comparing a textual description and a picture.

To accomplish this interesting task, we combined two approaches:

• Text mining—we extracted information from a product description and made a forecast as to what category this product might fall into;
• Image recognition—we analyzed a product picture to understand what could be shown.

Then, we compared the results of these approaches, determined what product we dealt with, and placed it into a corresponding category. While analyzing big data with our system, the accuracy of placing a product into the correct category was found to be way over 90% (almost 95-96%). However, when a human department performed the same task, the accuracy barely reached 80%. So, a huge assortment of products and an extremely routine job allowed for a greater number of human errors in comparison with our system. 

One more challenge—it was difficult to understand which item was on sale. For example, the company sold drapes, and we could have a picture of an apartment with a table, a couple of chairs, a picture on the wall, and some drapes in the window; the task was to single out the correct item for sale. To tell the truth, it wasn’t easy at all. 

It took us a long time to overcome the obstacle: we analyzed common elements in a group of pictures, compared them with textual elements, and so on. Our model was based on a probability estimation: for instance, it could determine that the item in question was a chair by 70%, a carpet by 60%, and a vase on the table by 90%. Being enforced by text analysis, our model offered a variant that was maximally close to reality.

As a result, our software allowed for:

1. Saving substantial sums on human resources because successful supervision of our system required just a few specialists to resolve disputable cases of maximally tricky recognition. 
2. Increasing sales considerably because an improved placing of goods into the right categories simplified their discovery by customers. If a company sells thousands of items, a categorization improvement by merely 1% means an increase in daily income by thousands of dollars.

We also introduced some minor improvements for warehouse management, integration with marketplaces, and others. The major approaches and technologies we employed during the project completion included e-commerce, artificial intelligence, data analysis, and optimization. 

3. Search System for Microelectronic Components

When Sonerim started its activities, we had a client with a product idea who was looking for a software development team to implement it. His project was backed by a prominent manufacturer and distributor of microelectronic components. Having worked in engineering and electronics manufacturing for a long time, he knew firsthand about the major problem in design engineering—the demand for components. So, we developed software for engineers to combat the problem. 

It isn’t common knowledge that manufacturers of basic electronic components (e.g., inductors, resistors, or microchips) are not numerous; it’s a highly monopolized market. If specialists need specific rare components for their work, the odds are good that they can be triply overpriced, and the alternative suppliers are few in number. 

The main idea of our project was the following—we developed a system that allowed for:

• Finding information about the availability and characteristics of different electronic components, 
• Comparing components’ features and discovering their best substitutes and analogs available in the market. 

Like Google, this search engine had a homepage with a search bar. We developed a rather advanced process: a client could write a very general, loosely-defined request for a certain component in the search bar, and the system could analyze it properly with its neat filters and come up with a suitable option. 

The selection of substitutes was the cherry on the cake. Let’s suppose an engineer had been employing one component for a long time, but a cessation of production or a significant price increase caused a necessity to find a substitute. It was difficult because only indicating voltage or resistance wasn’t enough; searching by parameters was a bad idea. Here are the reasons:

1. Not all data on all components were publicly accessible;
2. The publicly accessible data offered by a manufacturer didn’t always correspond to reality. 

Subtle electronic components behave differently in different batches because it’s difficult to manufacture components that reproduce the required parameters faithfully. Generally, they behave the way a manufacturer declares, but in detail, every batch can have deviations from the standard. Most of all, it refers to extreme temperatures: super-low and super-high temperatures can change characteristics to a significant extent. 

Our team dealt with the technical side of the project—the software development per se, while our client employed its data science engineers to create selection algorithms. We gathered data and analyzed them with the help of these algorithms in the following way: 

• First, when a new batch appeared in the market, we coded a few samples and carried out laboratory experiments to determine their actual behavior. 
• At the next stage, we developed models that allowed forecasting the behavior of all components in this batch.

Getting several samples from a certain batch manufactured at a certain plant by a certain shift was enough to predict the behavior of millions of other components. Moreover, these models empowered us to forecast the real parameters as opposed to those officially declared by manufacturers. We could undertake such a precise selection we had never seen in other systems. 

While visiting industry conferences, we reviewed a significant number of similar products. However, their developers merely compared the publicly accessible information from data sheets placed by manufacturers on their corporate websites. No subtle algorithmics was involved, unlike our product. Actually, any engineer could download the same experimentally unverified data from the web. 

Contrary to this approach, we offered real characteristics and built actual graphs based on the models we developed just for the purpose. In some cases, data provided by a manufacturer showed that a component could serve as a substitute, but our graphs indicated it didn’t match or only matched partially. For example, voltage can change a lot depending on the circumstances; extreme temperatures can cause deviations, and so on. If these parameters were irrelevant for a client, they could use the component; if variations were unacceptable, they could search further. In any case, it was an informed decision. 

Miscellaneous Projects

Scout is a disruptive application that makes the hiring process easy and convenient. You no longer need to waste time sorting through stacks of resumes, posting unfruitful listings, or paying recruiters outrageous commission prices. Scout will notify you about the best candidates and connect you with the top talent you want to see on your dream team. We’ll do the heavy lifting, and you conveniently pick from the best of the best. 

The technologies we employed during the project completion: 

• React.js
• PostgreSQL
• Node.js 

iCount is a leading online accounting service from Israel that provides solutions for online billing, invoicing, payment tracking, open APIs, and more. Currently used by large enterprises and small businesses, iCount attends to the needs of various companies that operate equally successfully with independent, self-employed professionals or thousands of employees. 

The technologies we employed during the project completion: 

• Amazon
• Objective-C
• Redis
• Java Android

SocialApps is a social network for sexual minorities. It allows finding new friends in a user’s location, saying ‘Hi’, chatting, and getting acquainted. While developing this software, special attention was paid to the stability of a video chat and the comfortable usage of the app. 

The technologies we employed during the project completion: 

• React.js
• Web-sockets
• Objective-C
• Node.js
• Twilio
• Java Android

Strong Authentication System is London’s Top 10 cybersecurity startup. It breaks through the ice of clumsy corporate monsters in the field of multifactor authentication, providing SaaS solutions with a flexible API for fast integration, modern auth tools like chatbots, and data signing to meet current challenges, reliable algorithms for OTP generation, and customer-oriented approach to make things happen easier. 

The technologies we employed during the project completion: 

• Java
• Tapestry
• Java Android
• Spring
• PostgreSQL
• Objective-C

Fisherman Navigation App allows finding your way to a selected destination, no matter your current location. The app is handy for tourists, fishermen, hunters, or scouts—anyone who prefers choosing their route. The support of different map formats isn’t a simple task by itself; however, the app’s primary goal is achieving good usability—you won’t read a heap of instructions or watch tutorials when you’re alone in the forest at night. 

The idea belongs to Sonerim’s teammate, Mychailo, a passionate fisherman who knows the ins and outs of this hobby. That is the reason why the app is loved by users so much—it has more than a million downloads on Google Play Market. Mychailo started developing an Android app on his own; a bit later, the company financially backed the development of the iOS version. 

The technologies our team employed during the project completion: 

• Java
• OSM
• Swift
• OZF

So, these are some of our projects. We like to tackle real-world problems and turn them into opportunities for our customers using the power of technology. At Sonerim, we are all about using data management and AI to rock the world. We have the skills, the tools, and the determination to take your company’s game to the next level. Let’s do it together!

No comments found for this post.