AI Superior

Summary: For a city municipality we developed a drone-based application that detects and reports 25 construction debris types. The solution allowed to considerably automatize construction site inspections process significantly reducing human involvement as well as reducing average inspection costs.

Challenge: A city municipality requested a solution to automate completed construction sites compliance monitoring to detect abandoned construction debris such as bricks, cement blocks, sand heaps, metal and wooden sticks, etc. It was crucial to have an automated solution that would minimize human involment to the inspection process thus reducing labor costs and time required for an inspection.

Solution by AI Superior: We applied our proprietary computer vision technology for object detection, classification, and segmentation to detect 25 different construction debris classes. We built a GIS dashboard to allow selection of a construction site and visualizing all the debris detected. Additionally, for every detected object the system provides an estimated size (area) of construction debris (for a single object and clusters of objects of the same type) as well as the amount of detected objects. The application provides insights be means of a GIS dashboard as well as exposes APIs to query detection results – this allows the solution to be integrated to practically any other systems.

Outcome and Implications: The solution was adopted by multiple city municipalities demonstrating its operational and economical effectiveness. According to customer estimates, the system saves 320 man-hours per month reducing average inspection costs by 40%.

The picture shows the result of the Construction debris detection from a drone. The different objects are dyed in different colors.

v\:* {behavior:url(#default#VML);}
o\:* {behavior:url(#default#VML);}
w\:* {behavior:url(#default#VML);}
.shape {behavior:url(#default#VML);}

Normal
0
false

21

false
false
false

DE
X-NONE
X-NONE

/* Style Definitions */
table.MsoNormalTable
{mso-style-name:"Normale Tabelle";
mso-tstyle-rowband-size:0;
mso-tstyle-colband-size:0;
mso-style-noshow:yes;
mso-style-priority:99;
mso-style-parent:"";
mso-padding-alt:0cm 5.4pt 0cm 5.4pt;
mso-para-margin-top:0cm;
mso-para-margin-right:0cm;
mso-para-margin-bottom:6.0pt;
mso-para-margin-left:0cm;
line-height:107%;
mso-pagination:widow-orphan;
font-size:11.0pt;
font-family:"Calibri",sans-serif;
mso-ansi-language:EN-US;
mso-fareast-language:EN-GB;}

Summary

We developed an AI component that translates a medical image from one domain to another using AI Style Transfer Technique.  The goal of the component is to enable re-use of existing software components and machine learning models that are employed to process a specific type of images i.e. images of tissues stained with one type of reagent across all other reagents.

Challenge

The customer required to find a way to apply the disease detection classifier trained on a specific domain of stained tissue images to other images of tissues stained by the different reagent. One of the biggest challenges of this task is the absence of paired images that allow one-to-one comparison e.g. black and white image to color image. Potential solutions include but are not limited to the creation of a new classifier and processing tool-set for each type of stain reagent. However, all these and other available alternatives are cost-intensive and time-consuming. Therefore, the customer decided to commission AI Superior to perform research and development activities to come up with the most appropriate solution for the task.

Solution by AI Superior

 AI Superior helped the customer to evaluate a set of state-of-the-art approaches applied to the histological images in a very short period. Among different learning methods, AI Superior also employed Generative Adversarial Networks (GAN) to do the unpaired image-to-image translation. GAN consists of a couple or more (depending on architecture) deep learning models. It focuses on excelling of image generation task that produces images visually similar to the input training set of images. In addition to the framework that allows translating images into the required domain AI Superior created an interactive visualization tool to validate the quality of generated images.

Summary: For a large semi-government organization we developed a system to detect litter objects from images captured from a drone as well as to manage trash collection activities. We designed and developed a GIS-based application allowing convenient interaction with detected litter objects as well as facilitating trash collection activities via optimal route planning and tracking of the collection progress.

Challenge: Many thousands of litter objects were nailed by waves to a coastal area scattering them across a territory of around one thousand square km. Being located on many single islands these objects were difficult and costly to identify by a human or a team of people. It was required to design and implement an efficient strategy to detect and collect those trash objects.

Solution by AI Superior: We applied our proprietary computer vision technology for object detection, classification, and segmentation as well as designed and developed an interactive GIS-based application to consume results and operate collection activities. The computer vision detection technology was applied to images captured from a drone. The drone flights missions were operated by a partnering team that was providing RGB images covering the whole territory. The flight altitude was around 50 meters high above the land.

Outcome and Implications: the developed solution allowed to reduce the time required to perform trash detection activities by the factor of 25. That resulted in significant cost savings halving the overall detection and collection costs. The collection time decreased by a factor of 4 while the automated computer vision solution allowed to achieve 7% higher detection accuracy compared to a human expert. Additionally, such a system decreased carbon footprint by the factor of 19.

The picture shows the dashboard of the application.

v\:* {behavior:url(#default#VML);}
o\:* {behavior:url(#default#VML);}
w\:* {behavior:url(#default#VML);}
.shape {behavior:url(#default#VML);}

Normal
0
false

21

false
false
false

DE
X-NONE
X-NONE

/* Style Definitions */
table.MsoNormalTable
{mso-style-name:"Normale Tabelle";
mso-tstyle-rowband-size:0;
mso-tstyle-colband-size:0;
mso-style-noshow:yes;
mso-style-priority:99;
mso-style-parent:"";
mso-padding-alt:0cm 5.4pt 0cm 5.4pt;
mso-para-margin-top:0cm;
mso-para-margin-right:0cm;
mso-para-margin-bottom:6.0pt;
mso-para-margin-left:0cm;
line-height:107%;
mso-pagination:widow-orphan;
font-size:11.0pt;
font-family:"Calibri",sans-serif;
mso-ansi-language:EN-US;
mso-fareast-language:EN-GB;}

Summary
We developed a road traffic analysis system for a system integrator executing a governmental project. The developed system can process video streams from traffic cameras and perform analysis of road scenes: detect objects and estimate their properties (trajectory, speed, etc.), identify traffic jams and violations, recognize car models and their license plates, and more. The system provides a safer environment for the city and the ability to monitor roads without human intervention.

Challenge

The customer, a system integrator executing governmental projects, required a road traffic analytics system that would identify different road entities (cars, pedestrians, bicycles, etc.), track them, and perform further analysis, e.g., identify car type, model, and make, color and other. The main chal-lenge was to create a robust system that could be easily transferable across different camera types and points of view, as well as to work in different environments.

Solution by AI Superior

The system that we developed includes multiple analytical components:

• Road entities detection (cars, pedestrians, bicycles, trucks, buses, motorcycles)
• Car type, brand, and model recognition (7000+ unique models)
• Object tracking
• Car color recognition
• License plate recognition and validation
• Illegal parking analysis
• Anomaly detection (e.g., a pedestrian on a highway)

Summary

For a banking organization, we developed an interactive tool that performs social media data analysis to facilitate marketing activities by understanding audience interests, social group affiliation, demographics, etc. The solution enabled the bank to provide discounts and offers to increase loyalty and decrease churn.

Challenge

The customer is a European bank serving hundreds of thousands of customers. To decrease churn and improve loyalty to its brand, the customer wanted us to develop a solution that would enable understanding of their clients, their interests, social status, etc. Based on this, the goal was to:

• Provide discounts and offers within a particular category, e.g., car credit or additional Visa card
• Provide offers from partners, e.g., cruises for those who like travel or sports equipment for people who enjoy sports
• Understand the audience for designing service packages, e.g., online service for young people
• Detect brand ambassadors in order to encourage them to promote new services

Solution by AI Superior

Please see additionally to the text our Video.

Based on customer requirements, we developed an analytical module and an interactive tool that allows us to extract and analyze insights obtained from social media data, visualize them and repre-sent them in a consumable form. Within this effort, we developed the following analytical modules:

• Audience interests extractor
• Social groups affiliation estimator
• The social class-based segmentation module
• Geography-based clustering module

Outcome and Implications

The developed Data Science Solution Tool  allowed the customer to interactively explore its audience and provide relevant offers and discounts that resulted in increased loyalty and decreased churn. According to customer feedback, the developed system allowed them to retain most clients that were previously going to churn.

Summary

To help a media company increase the lifetime value of their customers, we developed analytics to provide item recommendations from a diverse set of customer sources. The developed solution allowed the customer to increase the diversity of content consumed by their users by 5% and, as a result, increase the lifetime value of a customer.

Challenge

The customer is a big media company that owns various TV and radio channels, audio podcasts, magazines, and newspapers. They were interested in a personalized recommendation system for their existing users and content consumers. The content type is diverse (TV programs and shows, news articles, etc.). And every user has preferences that have to be understood and taken into consideration while recommending a new content item. The challenge is to create such a complex system that would recognize individual users’ consumption patterns, understand their content prefer-ences and recommend new content items that users are likely to consume. With such personalization capabilities, the customer is expected to increase engagement and decrease churn.

Solution by AI Superior

We developed a recommendation system that utilizes several factors to provide recommendations. The system has the following capabilities:

• Estimates consumption patterns of individual users
• Understands content preferences of each user (topics of interest, content type, etc.)
• Estimates demographics and technical means which are used by every user to access the con-tent
• Assesses content items similarity from different perspectives

To enable this, we developed several analytical components: NLP-based topic discovery and content tagging module, content items similarity extraction analytics, consumption patterns extractor, collaborative filtering-based recommender, item-to-item recommender, hybrid recommender that takes into account all the listed modules.

Summary

We developed automated candidate scoring and candidate-to-company matching systems that al-lowed our customer (an online recruiting company) to automate the process of assessing technical skills thoroughly. Due to the success of the solution, scoring experts were able to save 800 man-hours per month. Additionally, they enjoyed advanced functionality to their platform.

Challenge

The customer is an online recruiting platform that connects remote technical workforces with companies across the globe. Finding the right candidates with the right skill and expertise level is vital to providing quality candidate recommendations to companies looking to hire new workers. Additionally, automating the interview processes was crucial to scale up assessment procedures without making drastic changes to their internal human resources function. These two business challenges were reduced to technical ones:

• Automatic candidate scoring
• Automatic candidate-to-company fit estimation

Solution by AI Superior

To automate candidate scoring, we developed a Data Science System that utilized historical candidate data and the experts’ scores related to their performance during technical/coding interviews and tests. The Data Science system was trained to automatically score candidates based on their code, test outcomes, and technical challenges results. Additionally, to provide automatic candidate-to-company fit, we developed a Data Science solution that, by analyzing key features of companies and candidates, was able to provide the degree of the match according to various criteria. This allowed candidates to get the recommendation of companies to apply as well as provided candidates with an easy way to get the best fitting talent. This Data Science solution’s unique feature is the module that extracts skills from candidates’ profiles and enables skill matching and gap analysis in a fully automatic way.

Summary

For a niche insurance company operating in a medical/health domain we developed a prediction machine learning-based model to estimate the risk of an economical loss. The machine learning model is based on neural networks and built by consuming historical medical data over five consec-utive years. The developed model significantly outperformed statistical approaches. With this mod-el, the customer was able to optimize its pricing policies which resulted in significant savings.

Challenge

The insurance company operating in a medical/health domain was facing the challenge of pricing policies development. For them, it was important to understand risks related to a particular patient and adjust pricing policy models accordingly. In turn, the customer was expecting to experience considerable savings.

Solution by AI Superior

We built an application based on a machine learning model to predict the probabilities of a particular disease according to many input features and parameters including medical history. For that, we trained a deep learning model that was effectively dealing with intrinsic challenges such as class im-balance. Additionally, we built a validation framework to objectively compare multiple approaches and ensure that the created model was significantly outperforming others.

Outcome and Implications

The developed Data Science solution significantly outperformed the baseline models relying on statistics. The model outcome was used to optimize pricing policy to increase revenue and better manage risks.

Challenge

The customer is an international company with tens of thousands of employees interested in under-standing their employees’ skills. An important insight they wanted to obtain was to understand their business-critical employees’ skills and determine if a skill gap exists between those people and their potential successors.

Solution by AI Superior

We developed an interactive tool to perform a skills gap analysis between employees. The tool is based on NLP analytics that can:

• Extract and parse skills from unstructured text, e.g., position descriptions, performance re-views, CVs, etc.
• Bring these extracted skills to a three-level hierarchical skill profile structure and compare these skill profiles on different hierarchical levels.

Skills profiles are represented in a hierarchical structure that includes three levels:

• General level – This describes a high-level area of business activities, such as IT or finance.
• Function Level: This describes a particular function within the General Level group, such as software development (IT) or accounting (finance).
• Skill level: This includes specific skills extracted from employee profiles.

Outcome and Implications

The developed tool allowed the client to understand the skills of their employees, identify potential successors for business-critical roles, and notice which training is required to decrease identified skill gaps. The insights identified by the developed analytics helped to answer many other accompanying questions related to talent management, such as recruitment, succession, and training. The overall effort allowed the customer to save in total around 10,000 man-hours.

Summary
We developed a tool based on a machine learning model to automate real estate object estimation and find attractive offers on the market. The tool can also monitor the real estate market to identify ongoing trends. The component allowed the customer to obtain real estate market insights that are hidden from other stakeholders. Due to the tool’s success, the customer achieved a distinct competitive advantage and has already undertaken approximately 20 million euros worth of real estate investments based on the insights obtained through the solution.

Challenge

The customer is a real estate company operating in the DACH market and has the following objectives:

• to monitor and predict price dynamics on different property types (residential and commercial)
• to have an accurate automatic property evaluation
• to automatically discover attractive properties for investment purposes

Solution by AI Superior

We established a data-collection pipeline to analyze over one million real estate items and trained a deep learning model on that data. The resulting model allowed the customer to:

• Get a price estimation for a real estate object based on its parameters (house features, amount of land, availability of facilities, region and geographical location, etc.)
• Identify the degree with which each property parameter contributes to the overall price
• Predict market movement, capture temporal dynamics, and notify in case of a significant change
• Automatically identify underestimated properties on the market

Additionally, we developed an interactive visualization tool for user-friendly access to the model’s insights. 

Outcome and Implications

The system allowed the customer to react quickly to market changes and obtain high-quality insights for potential investments. Powered by the system’s insights, the customer performed around 20 million Euros worth of investments across multiple real estate properties that strongly demonstrated a positive return on investment.

Summary
Understanding the reasons and preventing customer churn is a critical component for the sustainability of the customer-oriented business. For one of the online gaming platforms, AI Superior created a machine learning model that learns player behaviour throughout the game and predicts the probability of churn for a particular time horizon. This solution helped a customer to predict the player’s churn event and employ the most relevant retention strategy. As a result of this project, our customer was able to decrease churn to 11,3%.

Challenge

Establishing a stable data collection and processing pipeline that didn’t affect the business operation due to high downtime was one of the most critical and challenging requirements from a customer. AI Superior’s data engineering team was able to successfully meet this requirement and build a high-quality data pipeline with almost no downtime of the customer’s online platform.

Solution by AI Superior

Machine Learning model that predicts churn over a time period for individual players. Moreover, model explainability has been integrated into the A/B testing framework for experimentation on retention strategies was also employed to test newly introduced features in the game.

Outcome and Implications:

allowed to understand churn reasons and developed retention strategies that reduce churn to 11.3%

Summary

We developed a service that extracts data from different types of invoices. Our service allowed the customer to automate the data entry process, ultimately saving 850 man-hours per month and decreasing the recognition error-rate by half. 

Challenge

The customer, a large multinational equipment distributor, was interested in streamlining their invoice processing activities through digitization. They wanted to extract account information from invoices in their custom-made ERP system. Prior to our involvement, the customer was doing this manually, and the process was taking several minutes per invoice. As the number of daily invoices began to grow, it soon became apparent that a human-centric solution was no longer feasible. Moreover, in their efforts to increase manual processing speed, more and more human errors were being made.

Solution by AI Superior

The solution delivered is a web service that allows users to upload an invoice and receive extracted information in a predefined structure. The service relies on a pipeline that consists of the following components:

• Document scans processing and enhancement.

• Invoice type classification.

• Identifying information blocks.

• Information extraction based on Optical Character Recognition (OCR) and post-processing.

• Information packaging and dumping.

Outcome and Implications

The developed service allowed the customer to streamline invoice processing in the CRM without any human intervention. It reduced the error rate by half since most errors fell into the category of human error. The speed of invoice processing increased by a factor of 25, and the customer saved 850 man-hours per month.

Summary

We developed a chatbot for a centralized meeting scheduling system. Our service allowed the customer to automate their process and reduce the number of human operators.

Challenge

The customer organizes meetings for his clients. Instead of clients communicating among themselves, communication takes place centrally through the customer. The customer was interested in automating this process. At the same time, the user communication with the system should be in natural language. Initially, the customer used Google DialogFlow. But the performance of Google’s solution did not allow the process to be fully automated.

Solution by AI Superior

We delivered a solution that receives response messages from the users to a meeting offer. Users can accept a meeting, decline it or propose an alternative date/time for the meeting. Moreover, the system can receive an automatic “out of office” reply. Our solution can extract the following information from a user’s message:

• The user’s intention for a meeting offer. The user could agree or refuse the meeting. He could suggest a new time or say that this time does not suit him and etc.
• If the user proposes a new meeting time or date our solution extracts a particular date and time from the raw text.
• If the system receives an “out of office” message our solution extracts the user’s return date if specified.

Summary

Availability of high-quality hyperspectral satellite imagery data unveils a new way of observing and monitoring the Earth. As part of our company's social responsibility activities, we developed an environmental monitoring solution based on satellite imagery analysis. The technology is based on an award-winning solution recognized by IEEE society. It allows us to identify objects on satellite images with very high accuracy. Detected objects can include but are not limited to: trees, cars, trains, buildings and identify coastal and sea debris, oil spillage, and other human activities results. 

Challenge

Detection of small (relative to the distance from where it was captured) objects on the Earth from an orbit is a non-trivial task. Environmental monitoring and object detection by humans using images of Earth from satellites is hardly achievable in the area of thousands of kilometers. In addition to that, some substances like methane leakage or oil spills might be invisible to a human eye which is sensitive only to three channels of the visual spectrum: red, green, and blue. 

Solution by AI Superior

We developed a Deep Learning solution that employs all channels provided by hyperspectral imagery and is able to recognize objects with a very low-resolution, e.g., 10 x 8 pixels. The approach demonstrates state-of-the-art performance in detecting various objects: residential and commercial buildings, cars, trains, roads, highways, railways, etc.

Outcome and Implications

The developed technology has a high potential to be applied in a wide variety of scenarios, namely:

• Ecology control and pollution monitoring: detection of coastal and marine plastic debris, oil spillage detection, methane leakage, detection, and severity analysis.

• Roads and railways quality monitoring, detection of road coverage by sand or snow, parking lots monitoring.

• Analysis of human activity on oil fields, in industrial zones, logistics monitoring.