Businessware Technologies

Client: A large pharmaceutical company with an extensive data management IT infrastructure. Objective: Develop a master data management system which would consolidate data from different sources into one application, remove duplicate data, and set up links between different sets of data. Results: We have developed a powerful master data management application for handling large amounts of data stored in business systems.  Use cases The system is designed to accommodate the most popular data management use cases like setting up new data directories and new entities within them, linking data from different sources together and setting up hierarchies between directories. Data filtering The MDM system features a robust data filtering system allowing users to set up complex dynamic data filters to easier find data and build reports.  Various Data Formats The MDM system we’ve developed is capable of storing multiple data formats, including text, dates, colors, hyperlinks, and links to other data entries. Access Rights The system supports user hierarchy and management of access rights to allow data editing only to authorized personnel. Data Export And Import The data can be imported and exported manually — a failsafe solution in case the connection between the MDM system and other applications is interrupted. Read More

Client: A digital transformation company focused on helping HR agencies automate their efforts with their workflow automation app, which includes handling of CVs, storing candidates' information and employer data, and more. Objective: Create an AI-powered CV processing system which would automatically extract relevant information and assess candidate's skill level. Results: We have developed an AI module for processing digital CVs. The module extracts important data, like candidate name, desired compensation, skills, experience, etc. in a matter of seconds irrespective of document formatting. ChatGPT For Entity Extraction Entity extraction, or detection of specific data from a document, is performed by ChatGPT 3.5. After the user sends a query, the app extracts entities from candidate CVs, prescribes them a qualification level, like "junior", "middle", or "senior" levels for software engineer position candidates based on their skills and achievements. Semantic Search We have implemented semantic search to enable users to input vague search queries and get specific results. After the entities are extracted, the app sorts through the relevant CVs and presents them to the user. We use Elasticsearch to sort through the candidate CVs and rank them according to how well they fit the search criteria. Read More

Client: A large company with hundreds of employees, multiple departments, and dozens of ongoing projects at any given time Objective: Develop a smart search system for the database which contains thousands of documents with project progress logs for the purpose of decreasing the time needed to search for information and restricting access to confidential or sensitive information based on the employee access level. Results: We have developed a smart chatbot powered by multiple language models which reduced the time employees spend looking for information by 35%. Multiple language models approach Given the confidential nature of the information stored in the database, using large language models like ChatGPT is not an option due to data leaking concerns. We have implemented a different approach to natural language processing by using a collection of multiple smaller language models to process data. Smart data search When a user asks a question, the system uses a retrieval language model which looks for parts of text which might contain the answer. After the texts have been located, a reranking model ranks them in order of relevance. The system can also combine text from multiple documents to provide a complete answer. Data access hierarchy  The system restricts access to sensitive information based on employee access level or department. Read More

Client:  A startup in the intelligent document processing field Objective: Create an ecosystem for insurance agencies to digitize their claim processing efforts capable of automatically detecting insurance claim type and layout, extracting typed and handwritten text. Results: We have created an AI-powered data extraction module for insurance claims that detects document structure and extracts relevant data in a matter of minutes. The module extracts data from the claims and prepares it for further processing. Detection Of Claim Type The module detects claim type (medical, business, etc.) and its layout, extracts claim type and title, and supports batch processing. Processing Of Multiple Template Types The AI module can work with various claim templates and designs, allowing for seamless data extraction from different claim types. The system can detect input fields and determine their type, like free-form text boxes, time and date fields, and more, as well as extract the data from the fields. Input Field Detection The AI module can detect multiple input field types, and extract data accordingly, for instance, it can detect checkboxes to determine which of the multiple options was marked. Detection Of Tables And Table Cells Claims our client works with often include tables with important data that needs to be extracted and processed. The module we have developed can analyze tables of any complexity, detect table cells, extract table titles and its content into an Excel file for further editing. Read More

Overview A secure app for secure communication between employees within a company. The messenger allows employees to exchange information without the fear of it being shared with a third party.  Challenge Our task was to create an iOS-based version of the app which would mirror the functionality of an already existing Android app. We had to adapt Android UI elements that are not native to iOS, as well as find a data encryption solution compatible with iOS. Solution The final app fully mirrored the functionality of the Android version while implementing data encryption completely differently.  To implement the generation of RSA keys, which is done out-of-the-box on Android, we have combined multiple libraries along with developing a custom formatting module which transforms the key signatures into those required by Android. We have also implemented secure push notifications. The notification itself does not contain the received message, but directions on how to find the message in a chat locally. The module replaces the notification received from the server with a secure push notification. Read More

Overview A modern and secure mobile app for a major bank which replaced an old version of the app and significantly improved user experience. Challenge We were tasked with the development of an Android-based banking app for a local bank which would replace an already existing outdated app. The new version had to include the following functionality: Log in and authorisation Bank account status and account management Manual and automatic money transfers Transaction history One of the primary goals was to not only develop an app which would update the design and improve user experience, but to ensure a high level of security as online banking often falls prey to hackers and malicious users. Solution We have implemented multi module app architecture which provides great benefits during the development process. Each module exists separately and can be implemented and tested individually. On top of that this approach makes the reuse of code easy and convenient. To ensure the safety of user data transfer, we have implemented end-to-end encryption to prevent data spoofing by malicious users through the replacement of a beneficiary account during money transfer. We used a androidx.crypto library to implement data encryption as an out-of-the-box solution was the most cost efficient option for the client. Read More

 Overview A cross-platform app for monitoring a 3D printing process remotely. The app automatically detects printing malfunctions and alerts the user when the printing process is complete. Challenge We were approached to create a mobile app which would mirror the functionality of an existing website - a SaaS video streaming service to help monitor the printing process in real time, allowing anyone with a 3D printer to remotely observe the operations.  Solution We have developed a cross-platform app based on WebRTC and Janus Gateway. The app not only shows the stream, but also includes all of the relevant information available about the process - time to completion, printer temperature, printer status, etc. The app has the following features: Social login - users can log in through their Facebook and Google accounts Status updates through push notifications - users get updates on the printing process through push notifications that show the progress bar and a stream screenshot Status monitoring - the app shows data on the printer’s temperature, printing progress, estimated completion time, printer status and more The app is augmented with artificial intelligence that automatically detects various malfunctions, like spaghetti or model dislocation, and sends notifications to the user’s phone to let them know something might have gone wrong. Read More

Overview Our client decided to take advantage of the latest developments in natural language processing and has come to us with an interesting project aimed at automating the market analysis. Our goal was to develop a system that would determine customer pain points and determine what people like and don’t like about similar products already on the market. Challenge The system had to receive a website page describing the product in question as input, determine what the product was, gather the public’s opinion on similar products and determine main advantages and disadvantages according to the customers. While working on this project, we had to address the following challenges: The system had to successfully determine what the product in question is We had to determine where the system would go to gather the public’s opinion The system had to integrate GPT-3 and produce human-like written content Solution First, the system receives a landing page with the product’s description where it collects keywords to determine what the product is and its main features. The system looks for reviews to similar products on Amazon and pulls them out, after which we access GPT-3 via an API to generate a text detailing the main advantages and disadvantages of the rivaling products. This information can then be used as the basis for any marketing content or as marketing content in itself in a form of a script for a promotional video or a social media post. Read More

Overview As more and more media content - like video games, videos, films and cartoons - is created, it gets harder for production companies and investors to decide which content is worth promoting and investing capital into. There is a strong incentive to heavily filter through the content from creators on the basis of potential commercial success. While potential success is difficult to measure and is often determined based on personal experience, it can be formalised to predict how popular the piece of media in question will be with consumers. By analysing different aspects of already popular media, one can learn what determines the success of said media with a high degree of confidence. Challenge Our client is a production company which helps creators promote media content, like mobile and browser video games, attract more customers and generally increase its popularity. Our client helps hundreds of independent creators and media development studios, working with dozens of new projects daily, providing consultations and giving recommendations on how their products can be improved. In order to give more accurate recommendations and determine which projects will be successful in the future, our client asked us to develop a system for analysis of digital media, like a video game recording. Solution We first determined the video framerate which would yield satisfactory results while maintaining quick processing speed. Each frame is being quantized by 16 colors, meaning all colors are divided into 16 groups, which is the most optimal in terms of processing time and visual perception. The system then detects top 8 dominant colors present in the video. After the dominant colors are determined, the application compares the color specters to popular color harmonies to provide additional information for our client. Depending on how much the media in question adheres to a certain color harmony, recommendations on possible design changes can be given based on information on previous projects.  The system also detects scene changes via monitoring the color histogram and detecting sudden and drastic changes in dominant colors or a sudden appearance of one or more new colors, as well as calculates the camera speed by locating static objects and building a motion vector.  Read More

Overview With the use of machine learning and NLP approach, we have developed a system capable of producing human-like social media captions in seconds. Challenge We were tasked with the development of a system which would take an image as input and provide a social media caption as output. The text had to not only match the content of the image, but also be indistinguishable from that written by a human. The ‘human-like’ aspect of the generated texts was of the utmost importance - users can easily identify AI generated text which reduces the engagement and customer trust. Solution The process of generating a text based on an image consists of two stages: Object recognition stage - the system has to detect objects present in the image and identify them Text generation stage - based on the identified objects, the system has to generate a human-like caption suited for a social media post The model which we have developed generates social media captions which are indistinguishable from those written by real users in tone and grammatical structure. The model embellishes the generated text with appropriate emojis and hashtags to further imitate what an average social media user would write under an image Read More

Overview First Step AI is a startup looking to bring machine learning to the masses. The main goal of this project is to create a system that would be simple for the average user to train their own image recognition model, but complex enough to fit the needs of software developers and machine learning specialists. We have developed a machine learning model training platform that allows users to create, train and test their own ML models. The platform consists of several modules which reflect the user's journey. Projects All models trained using the platform exist in the form of projects - a collection of all relevant data, allowing for a quick overview of all user projects and their status. Dataset The platform allows users to add and annotate their own datasets by uploading images and marking them up. Before the markup starts, users can create labels that represent the objects which need to be detected to use for a markup later. After the images are ready, the platform performs dataset analysis and provides relevant statistical information, like total image count, along with a balancing score that shows how good the dataset is in terms of objects distribution by classes. Image Recognition Model Training After the dataset is complete, users can start training their own machine learning model. They can choose Tensorflow or PyTorch with various sets of base models and input sizes. After the model has been trained, users can check on the final mAp and decide on whether it is high enough or if they want to improve it.  Model Testing And Download If the users are happy with the mAp, they can test their models within our platform by uploading images and videos not included in the training dataset and seeing how well the model is able to detect the objects. Once the testing is complete, users can download the model in several formats depending on how it will be used. 3D Model Generation An additional feature, 3D model generation allows users to upload photos of any object and get a detailed 3D model, which can then be used for 3D printing. Read More

Overview We have developed a computer vision monitoring system for a bottling factory. The system helps to identify improperly canned bottles and low filling lines, reducing the number of defective bottles and improving customer satisfaction by providing consistent quality of products. The system is more efficient than daily checkups of random bottles as it detects every faulty bottle, eliminating the need to throw out a whole batch. Challenge Our client is a beverage bottling facility, processing hundreds of tin and glass bottles daily. Bottling line sampling, or hourly inspections of randomly picked bottles, is not effective nor efficient. It takes a significant amount of time to pick the bottles among dozens of containers, assess fill height, label placement, cork height, etc. Faulty bottles can easily slip by the inspector's eye, decreasing user satisfaction and resulting in a monetary loss for the company. Solution The facility was equipped with a set of three security cameras which were installed at several stations along the conveyor belt and placed in a way that provided a clear view of the bottles. The ML model trained on the dataset is capable of detecting various events at different stations in real-time. The system monitors three conveyor belt stations - filler, canning, and labeling: Filler Station Filling lines - how much liquid is in a bottle - can be either too low or too high, depending on the nature of a mishap: a deformed bottle, degraded water seal, etc. The system can detect when a bottle is not filled properly and alert the personnel. Canning Station The canning process consists of two parts: a cap is placed on top of a bottle, after which it is pressed on to seal the bottle. Sometimes, due to liquid overflow, a deformed bottle, or equipment malfunction, the cap is not placed in a correct way or not placed at all, resulting in an unsealed bottle which affects the bottling process down the line - wetting down the labels making them soggy and therefore unreadable, or affect the beverage shelf life, resulting is a health risk for consumers. Our system detects if the cap has been placed correctly and if the bottle is sealed properly. Predictive Maintenance The system can not only be used to detect faulty bottles but to assess the condition of equipment, detecting early signs of a breakdown. For example, bottles are usually filled low due to a degraded or broken water seal, which can lead to a complete production shutdown, causing significant monetary losses. Our system can detect these issues early on and let the client know the equipment needs to be inspected Read More

Overview In the age of product marketplaces and steep competition, keeping an eye out for customer feedback is essential for marketing products. We were approached by a marketing agency looking to gather customer reviews of competitors' products on Amazon and analyze them to create better marketing materials with customer feedback in mind. Challenge Our client has approached us to create a system that would analyze Amazon reviews and show the customers' feedback about the product in a form of product aspects, i.e. customer opinions grouped by topic: what aspects have they found to be the most satisfying about the product what disappointed them the most about the product aspects they have found surprising, interesting, etc. Solution The system we have developed uses Amazon API to extract Amazon reviews for any given product. After the reviews are extracted, the system splits reviews into sentences, groups them by topic, and chooses a sentence that best describes the topic. Results are presented in a form of aspects - main themes extracted from reviews. Each aspect can be expanded to see sentences from which the aspect was extracted. Sentence processing is done using multiple NLP and machine learning models using GPU for training. The system can correctly process natural language, including slang, abbreviations, and contractions. Read More

Overview A digital guide for all museum-goers, our image recognition app is able to detect art pieces from any angle instantly and provide information about the piece in text form. Challenge The app was created for an art museum looking to provide digital guide services to their visitors through their smartphones: the visitor would have to point their smartphone camera at an art piece, and its description would show up on the screen. The project had significant limitations: Ability to add more classes without retraining the model - we had to create an app that would allow the addition of new classes (new art pieces) without the involvement of ML developers Recognition speed - our client requested the recognition process to take less than 1 second Solution The app uses keypoint detection to recognize art pieces in real-time and provide their description. The algorithm works just as well as a machine learning model would for their application, but meets the requirements of the project, which regular ML models couldn’t do. This app is a simple and elegant solution for those who want to reap the benefits of AI, but don’t - or can’t - get deep into it, investing money not only into the app development itself but the retraining and general maintenance that comes with AI models. The app’s recognition module works without any supervision, our client can add new art pieces as often as they please. Read More

Overview A PDF recognition system for feature extraction from engineering drawings with AI. The system reads engineering drawings and returns structured information, including the total count and type of doors, windows, walls, etc. Challenge To accurately prepare a bill of quantities and provide a price estimation, the following information needs to be extracted from a document: Building type - office building, shopping mall, apartment building Floor plan type - electrical, plumbing, etc. Number of doors, windows, and rooms Total wall length Total area The documents usually contain multiple technical drawings, especially when it comes to multi-story buildings.  Ready-made solutions for optical character recognition (OCR) could not handle the task with enough accuracy or could not handle it at all due to the special characters used in the technical drawings, as well as the graphic nature of the drawings themselves. Solution The first step in analyzing any PDF file with a floor plan is to detect the location of the floor plan within the page. We have developed a segmentation machine learning model which automatically detects the drawing location. The system also gives users the ability to highlight the floor plan themselves. The system detects technical drawing type and scale, and automatically generates a table of contents, making it easy to navigate large multi-page documents.  Object Detection The application automatically detects walls, doors, windows, electrical outlets and other objects, counting them and detecting their types by automatically reading the labels. The data can be extracted in a form of an Excel table. PDF table to an Excel table Floor plans come with a bill of quantities that contains information about the different labels used in the technical drawing, as well as information about the materials. The PDF tables are not an ideal way to handle large amounts of data since they cannot be edited and the data cannot be sorted or filtered. There is a number of readymade tools and solutions that can turn a PDF table into an Excel one, though they work poorly with large complex tables that include merged cells and span across multiple PDF pages. Readymade solutions do not handle merged cells well and often split them incorrectly Often when PDF tables span across multiple pages, their columns don’t line up which causes existing tools to process the data incorrectly If the text goes outside of its cell, readymade solutions split the text into multiple cells. We have developed a subsystem that scans the PDF tables and turns them into Excel tables without changing the original structure of the table and keeping the data integrity. Read More

Augmented Reality App for Metropolitan Museum of Art

Meet the Augmented Reality (AR) application, introducing an interactive map for the Metropolitan Museum of Art (MET). Avoid getting lost in one of the most famous museums in the world with our easy-to-use application that allows you to locate any of the galleries, showrooms and other public places on the go! Read More