<h3>Solution &amp; objective</h3><ul><li>The main objective with our classroom format AI/ML Fundamentals course is to remove the hype and the black box around AI and ML (Machine Learning).</li><li>One course session is 6-7h long with a mix of theory and hands-on training in Google Colab. Participants get to build and trains their own ML models, no programming skills as prerequisite.</li><li>Course material is comprehensive but well-structured and provided by highly pedagogical and practically experienced tutors.</li></ul><h3>Value</h3><p>After finalizing the course, attendees revealed that they:​</p><ul><li>Were able to understand AI and ML beyond the hype and felt much more confident and prepared for the AI related decisions and actions on their ongoing AI/ML journeys, with business benefits as outcomes.​</li><li>Could better understand data scientists and data engineers in their teams/projects and expected more efficient collaboration with them.​</li><li>Were aware of the hidden costs in the AI projects ,understood the importance of having good quality data and gained proper knowledge about the AI/ML workflow and what it takes to train and build decision forecasting machines, accurately.</li></ul>

We provided one day AI/ML Fundamentals training course for Get AI members, among them employees from Öresundskraft and Helsingborg Stad.

<h2>Challenge</h2><p>Placing wind and solar power plants in the right geographical location requires exploration of available land with the right characteristics. This kind of site selection is typically performed manually with rule-based filtering in GIS (Geographic Information System) data and this task is rather time-consuming.</p><h2>Solution</h2><p>A dump of GIS data was analysed focusing on the pixelated (or gridded) raster/slope data where each pixel is associated with a specific geographical location. With insights from analysed data, we implemented and delivered AI-optimized parameters for more optimal use in the rule-based filtering.</p><h2>Value</h2><p>Analysis and filtering of the GIS-data with AI-optimized parameters instead of rule-based is more efficient, i.e. less time-consuming.</p>

Implementation of AI-optimized parameters for more optimal placing of wind and solar power plants in the right geographical location.

<figure class="table"><table><tbody><tr><td><h2>Challenges</h2><p>LF Skåne insures around 5,000 farm properties in Skåne, mostly former farms now used as residences and for various activities. These properties typically have multiple detached buildings, such as houses, barns, and stables, which can serve multiple functions. For example, an old barn might house a stable, a machine hall, and an art studio. This requires a unique type of insurance with many different insured items per property, often exceeding the number of buildings. Manually inspecting these properties is a time-consuming task for LF Skåne employees.</p><h2>Solution</h2><p>Various APIs offering satellite images, aerial photos, and maps were explored for use in AI/ML models to facilitate AI-driven digital analysis of farm properties in Skåne. Challenges include a lack of baseline truth in insurance data, missing plot boundary data, difficulties with correct address identification by map services like Google Maps, and the limitations of semantic segmentation models trained on ground-level images rather than satellite or aerial images.</p><p>A prototype digital tool was developed for analysing properties using suitable API data, aiming to identify discrepancies from insurance policies. Specifically, the project scope included using digital/AI analysis to investigate counting buildings and compare them to insurance data, identify building types, detect solar panels, and assess roof conditions to identify "bad" roofs.</p><h2>Value</h2><p>A digital farm property analysing tool can be used by insurance investigators for preparations upon farm property visits reducing visit time per farm and potentially also identifying discrepancies factors investigators might miss during their physical investigations.</p></td></tr></tbody></table></figure>

Exploration of various APIs providing satellite images, aerial photos, and creation of digital prototype tool for digital analyses and investigation of farm properties in Skåne.

<h3>Challenges</h3><p>Alfa Laval being a big organisation and this project their first AI/ML project initiative. Time and internal resources needed for getting approvals and extracting data from various data sources.​ Low predictive power in Alfa Lavals customer data.&nbsp;</p><p>Used dataset contained too few examples of confirmed churned customers along with huge diversity in customers and customer behavior. ​</p><h3>Solution</h3><ul><li>PoC solution with churn scoring model looking at time series data from customers that already have churned by investigating their characteristics / behavior / predictors such as volume, frequency, regularity, purchased products, before the churn.​</li><li>Work was done in Microsoft Azure platform in combination with open source tools and frameworks. Power BI was used for the visualization of solution. Agile way of delivery, with 2-weeks sprints.​</li></ul><h3>Value</h3><ul><li>Revenue lost from churning customers has for used dataset been calculated to over 116.000 EUR daily and near 10.5 M EUR quarterly.​</li><li>Project detected bad quality on Alfa Laval's customer related data which raised greater data strategy, management and governance initiative in the organization with mission that more accurate and complete data is entered and saved in the source systems and their databases going forward.​</li></ul><figure class="image"><img className='content-img'  src="https://firebasestorage.googleapis.com/v0/b/nordaxon-webapp.appspot.com/o/Alfa%20Laval%20-%20Customer%20Churn%20Prediction.png.jpg?alt=media&amp;token=aa2ec91b-f1c6-4878-8dc0-300c43486f48"></figure>

Explored eight years of Alfa Laval’s own customers’ purchase behaviour data and built a PoC model predicting when customer churns i.e. stops doing business with Alfa Laval. Throughout project execution explained frameworks used and steps in a typical supervised ML workflow applied.​

<p><strong>Challenge</strong><br>LindInvent provides sensor-driven regulatory products that intelligently control indoor climate, lighting, ventilation, and sun shading in smart buildings based on real-time demand. Configuring and optimizing these regulators requires adjusting numerous parameters, which can lead to anomaly deviations and errors that can be difficult to detect and resolve, for example variations in airflow and temperature. NordAxon was assigned to explore AI-driven methods to improve anomaly detection and system efficiency.</p><p><strong>Solution</strong><br>The project focused on applying machine learning techniques to detect and cluster anomalies in a subset of LindInvets indoor climate data. Initially, conventional anomaly detection methods were tested, followed by a deep-learning approach using LSTM-based autoencoders. After evaluating different strategies, the project refined its approach to use unsupervised clustering for anomaly detection in air diffusers, leveraging structured data to extract insights.</p><p><strong>Value</strong><br>Unsupervised anomaly detection was identified as the most viable approach due to the lack of labelled data, though it cannot fully explain anomalies.&nbsp; Predicting system failures with data within the scope of the project was found to be infeasible due to missing labels and high complexity in system algorithms. As the causes of the anomalies are more interesting than the actual anomaly further work will focus on exploring the regulatory dependencies and create smart alarms with explanation of the anomaly.&nbsp;</p>

Exploratory anomaly detection in LindInvent’s smart building systems using unsupervised clustering to identify airflow and temperature deviations, enabling smarter alarm solutions.

<h2>Challenges</h2><p>In big organizations internal documents accumulate over time&nbsp;making it challenging for employees to find the necessary information quickly.</p><h2>Solution</h2><p>Within the scope of prototype implementation, necessary data&nbsp;was retrieved from various data sources (SharePoint, e-mail boxes and external web pages) and stored in a vector database. Number of large-scale language models (LLMs), both open och closed, were investigated and best suitable for the project scope were&nbsp;prompted and fine-tuned with&nbsp;the&nbsp;retrieved data together along with the power of RegEx.&nbsp;For interaction with the models, a simpler StreamLit interface was developed.</p><h2>Value</h2><p>Employees often&nbsp;spend time on&nbsp;manual, time-consuming tasks that can be automated&nbsp;and replaced&nbsp;with today's AI/ML-driven technology.By allowing employees to chat with internal documents, external and internal pages, and mailboxes, it is possible for them to get the information they are looking for quickly.&nbsp;Using NLP/LLM&nbsp;technology for summarizing,&nbsp;generating,&nbsp;and looking up in text&nbsp;is a time saver.</p>

Developed LLM-driven prototypes chatting with the company's internal data and external web pages.

<h3>Solution &amp; objective</h3><ul><li>The main objective with our classroom format AI/ML Fundamentals course is to remove the hype and the black box around AI and ML (Machine Learning).</li><li>One course session is 6-9 long with a mix of theory and hands-on training in Google Colab. Participants get to build and trains their own ML models, no programming skills as prerequisite.​</li><li>Course material is comprehensive but well-structured and provided by highly pedagogical and practically experienced tutors.​</li></ul><p>&nbsp;</p><h3>Value</h3><p>After finalizing the course, attendees revealed that they:​</p><ul><li>Were able to understand AI and ML beyond the hype and felt much more confident and prepared for the AI related decisions and actions on their ongoing AI/ML journeys, with business benefits as outcomes.​</li><li>Could better understand data scientists and data engineers in their teams/projects and expected more efficient collaboration with them.​</li><li>Were aware of the hidden costs in the AI projects, understood the importance of having good quality data and gained proper knowledge about the AI/ML workflow and what it takes to train and build decision forecasting machines, accurately.​</li></ul>

During three separate full day course sessions we provided AI/ML Fundamentals training course for approximately 35 IT leaders and decision makers.​

<p><strong>Challenge</strong><br>Tetra Pak, a global leader in packaging solutions, continuously strives to optimize machine performance and minimize downtime in its manufacturing operations. To tackle this, they wanted NordAxon to explore AI-driven methods for predicting machine stops and extracting actionable and valuable insights from internal production data, with a particular focus on their A3 CompactFlex filling machine.</p><p><strong>Solution</strong><br>NordAxon developed and tested multiple AI models and methods, including pattern-mining, BERT-based classification, and multiclass classification approaches. The models analysed event sequences to identify predictive patterns, but performance was hindered by the high-level nature of the logged data. Despite the challenges, valuable learnings emerged about the limitations and requirements for future AI-driven anomaly detection.</p><p><strong>Value</strong><br>The project provided critical insights into the feasibility of AI in predictive maintenance for Tetra Pak's production lines. It highlighted the need for more granular sensor data and better integration of performance, sales, and technical issue records for the given problem statement. These findings will guide future AI initiatives, optimizing machine performance and reducing downtime in manufacturing operations.</p>

Exploratory predictive maintenance of filling machine performance data, aiming to forecast machine stops and generate actionable insights

<p><strong>Challenge</strong><br>Trackfellow is an app built for dog handlers who love tracking—whether it’s blood tracking, mantrailing, or competition trails. The goal was to create smart, AI-driven features that automatically generate training tracks and analyze performance over time to reveal what each dog needs to improve. The main challenge was the niche nature of the domain, where no standard off-the-shelf AI solutions exist, and the need to seamlessly integrate the functionality into a simple, user-friendly app.</p><p><strong>Solution</strong><br>NordAxon developed an adaptive solution combining a rule-based algorithm with FP-growth, an advanced AI/ML technique. This hybrid approach allows the system to learn from each dog’s performance, pinpointing which factors influence track difficulty. The final solution was delivered as an API, making it simple to integrate with the Trackfellow app and extend with future AI capabilities.</p><p><strong>Value</strong><br>The result from this solution is a smarter, more personalized training experience. Dog handlers gain meaningful insights into their dogs’ progress, while Trackfellow becomes a powerful, AI-driven companion that helps users plan, train, and improve more effectively.</p>

AI-powered training in Trackfellow’s dog tracking app that learns from each dog’s performance to create personalized tracks and make every training session smarter and more effective.

<h4>Challenges</h4><p>One challenge with traditional survey methods is a lack of engagement from participants. Static forms and monotonous questionnaires can lead to low response rates and incomplete answers. This disengagement results in a barrier for collecting answers, as participants may find the process tedious or irrelevant to their experiences.</p><h4>Solution</h4><p>To overcome this challenge, an interactive chat-based form with generative AI capabilities was developed. This solution offers a more natural and engaging interface, akin to a conversation, making it easier and more appealing for participants to complete surveys. The generative AI not only collects responses but can also interact dynamically with participants, asking follow-up questions and clarifying responses in real-time. This approach simulates a natural conversation, thereby significantly enhancing user engagement and the quality of the data collected. In addition, generative AI was utilised for generating PowerPoint presentations based on the answers, allowing for faster analysis and decisions.&nbsp;</p><h4>Value</h4><p>The interactive nature of this chat-based form significantly increases participant engagement and the completion rate of surveys. By providing a more user-friendly and conversational interface, it lowers the barrier to completing surveys, encouraging more thorough and thoughtful responses. This leads to richer, more reliable data, allowing organizations to gain deeper insights into participant feedback. The use of generative AI in analyzing responses further adds value by offering nuanced and comprehensive analysis, aiding in informed decision-making.</p>

Developed an interactive solution for collecting surveys, with generative AI capabilities for analysing and transforming gathered answers.

<p><strong>Challenge</strong><br><span style="color:hsl(0,0%,0%);">A global leader in document identity verification, turned to NordAxon to overcome the challenge of automating the classification of over 2,000 types of identity documents with both speed and precision. The challenge was to automatically identify more than 2,000 document types within strict latency limits while remaining CPU-efficient for large-scale deployment.</span></p><p><strong>Solution</strong><br>NordAxon team of ML engineers engineered and implemented an optimized ML-driven image classification pipeline designed for high throughput and precision. By integrating model conversion for lightweight deployment, inference optimization for latency reduction, and vector search for rapid similarity matching, the solution delivers top-ranked predictions within 300 milliseconds — maintaining exceptional efficiency even on standard CPU environments.</p><p><strong>Value</strong><br>NordAxon’s solution achieved remarkable results — 99.6% accuracy among the top 10 predictions and an average processing time for each document in only 90 milliseconds. With this capability, our client can perform near-instant document verification at scale, while remaining ready for future advancements like GPU acceleration and live data integration.</p>

High-accuracy, low-latency AI image classification for automated recognition of diverse identity documents at scale.

<h3>Challenges</h3><ul><li>Many dependencies between different workstreams and projects in the AML Programme as whole, caused delivery delays.</li><li>Sourcing data needed in the flow coming from many different systems and governance models.​</li><li>Deciding on which data is critical for the flow and which controls on that data are necessary.​</li><li>FSA (Financial Security Authority) inspections from local branches interpreting project work.​</li></ul><h3>Solution</h3><ul><li>The bank´s AML TM project ensures automated, integrated and controlled flow for all payment transactions entering the bank externally or being generated internally. Data in the flow is sourced from many systems before being monitored and alerted by SAS AML system. ​</li><li>Data remediation workstream is responsible for delivering relevant, enough and data to the TM flow by selecting critical data elements and implementing data quality and measurement controls on those.</li></ul><h3>Value</h3><ul><li>The bank's money laundering scandal arose in 2017-2018 around €200 billion of suspicious transactions that flowed through the Estonia-based bank branch of the bank from 2007 to 2015.​</li><li>Fines of several billion dollars have been payed to financial regulators in Denmark, US and other European countries and upon new breaches more fees will be required.​</li></ul>

Acted as compliance SME (Subject Matter Expert) in Data Remediation workstream for TM (Transaction Monitoring) project within AML (Anti Money Laundering) Programme. Responsibilities covered data management, data quality and advanced analytics questions.​

<p><strong>Challenge</strong><br>Schools might struggle with finding students who are at a risk of not attaining base qualification for high school in time. Finding these students early would give greater possibilities to focus resources where they are needed the most. The schools also need insights into which factors affect student grade performance the most on a school-level, enabling more efficient resource-allocation across the board.&nbsp;</p><p><strong>Solution</strong><br>Using several data sources, an explorative analysis was performed. Take-aways from the analysis were used in creation of several machine-learning experiments, evaluating the feasibility of predicting student qualifications based on early student-specific data. The analysis is also used to generate insights for relevant stakeholder.&nbsp;</p><p><strong>Value</strong><br>Creation of an accurate model would mean greater possibilities for schools to focus efforts where needed. Analysis gave new insights and reinforced previous conceptions of what factors correlate with student performance.&nbsp;</p>

Exploring the feasibility of predicting high-school qualification using early data on socioeconomic factors, student absence, and performance through AI-driven analysis.

<h3>Solution &amp; Objective</h3><ul><li>The main objective was to learn the students the fundamentals of Statistics and how this is applied in real world practices.</li><li>Over the course of four weeks the students where to learn the fundamentals of Statistics, such as maths, descriptive statistics, Bayes theory, confidence interval, binomial distribution etc.​</li><li>Course material is comprehensive but well-structured and provided by highly pedagogical and practically experienced tutors.​</li></ul><p>&nbsp;</p><h3>Value</h3><p>After finalizing the course, the students had:​</p><ul><li>Gotten an overview of statistics and how to use it.​</li><li>Got the opportunity to ask critical questions regarding statistical concepts.​</li><li>Learned to create statistical calculations to better prepare for professional life.​</li><li>Got an introduction to programming in Python.</li></ul>

As a part of the IHM Business Schools Digital Analytics Specialist education teaching approximately 40 students a course in Statistics. Assignment included preparing and holding lectures, practical assignments and a graduate test at the end of the course. ​

<h3>Challenges</h3><ul><li>Many parts of the organization had limited understanding of AI/ML, making their needs and requirements hard to communicate.​</li><li>Objective is to develop WebGUI for tuning cameras when using ML object detection system for optimal inference which requires research and development of an objective image quality metric.​</li><li>Due to lack of complete dataset for Swedish traffic sign detection, the solution required a combination of mixed datasets and an unconventional solution when using them.​</li></ul><p>&nbsp;</p><h3>Solution</h3><ul><li>A single-shot object detection model that reliably detects Swedish traffic signs and measures image quality by number of detections/accuracy, was developed and deployed on the sensor edge.​</li><li>Semi-supervised model was trained and built by fitting YOLOv3 algorithm into two image datasets; Linköpings Traffic Signs (2 500 labeled and 15 000 unlabeled) and COCO 2017 (100 000 labeled).​</li><li>Components used in the infrastructure architecture: a deserializer board and camera modules , NVIDIA Jetson AGX Xavier GPU.​</li></ul><p>&nbsp;</p><h3>Value</h3><ul><li>AI processing systems built into the hardware can discard all kinds of unnecessary or unwanted data before that data ever goes into the main device’s storage or processing pipeline. This decreases processor power and removes latency making object detection safer and more secure.​</li><li>ML insights obtained during the model training and testing deliveries of great value for hardware development.​</li><li>The new detection model is proven to have market competitive advantages for the company.​</li></ul><figure class="image"><img className='content-img'  src="https://firebasestorage.googleapis.com/v0/b/nordaxon-webapp.appspot.com/o/Sony%20-%20Inne%20i%20case%20texten%20-%20Edge%20AI.png?alt=media&amp;token=0d6bce42-5a78-4eee-9dc0-fee77794e899"></figure>

Improved ML models for Edge AI camera systems and enhanced team expertise in modern ML frameworks.

<h2>Challenges</h2><ul><li>Removing the hype surrounding the concepts of AI and ML by explaining the technology behind them in an intuitive way.</li><li>Explaining the history leading up to today's AI and ML capability, both in terms of possibilities and limitations.</li><li>Giving leaders and decision makers a technical foundation to help make informed and proactive decisions in AI-driven projects.</li></ul><h2>Solution</h2><ul><li>A one day course, presenting everything from the history behind AI, to the distinction between classical machine learning and modern deep learning.</li><li>Presenting common pit-falls and challenges within AI/ML projects, and how to combat them in a strategic and operational level.</li><li>The lecturers are experienced tutors from NordAxon and partners from academia.</li><li>A mix of theory and practice, with no need for prior programming experience.</li></ul><h2>Value</h2><p>After one full day session, the course review with the participants showed that:</p><ul><li>All participants would recommend the course to others.</li><li>80% of the participants felt more ready to work with AI.</li><li>When asked what they learned, some of the responses were:<ul><li>&nbsp;“<i>More insight in what the different types of models can and cannot do. </i>”&nbsp;</li><li>“<i>Clarity in the meaning of words and connections - greater understanding of the complexity.&nbsp;</i>”</li><li>“<i>That we need a data governance framework and need to look more into CDO. </i>”&nbsp;</li><li>“<i>The importance of monitoring models. </i>”&nbsp;</li></ul></li></ul>

Our popular one-day crash course in AI/ML fundamentals was requested and appreciated by the management team at the Øresund bridge.

<h4>Challenges</h4><ul><li>Quickly understanding a variety of different businesses and their needs, and relating them to existing AI/ML and data possibilities.</li><li>Explaining some of the latest ML break troughs to potentially non-technical business owners.</li></ul><p>&nbsp;</p><h4>Solution</h4><ul><li>Guiding startups and scaleups in the AI/ML field and how they can work in a data-driven manner.</li><li>Monthly AI-related events and educational presentations for Hetch's incubator and accelerator programs.</li><li>Regular “open office” hours for any and all AI-related questions, thus acting as both advisors and educators within the field.</li></ul><p>&nbsp;</p><h4>Value</h4><ul><li>Supplementing Hetch's offerings and programs with expertise in AI and data science.</li><li>Advising startups and scaleups in their AI-journey and in how to work in a data-driven manner.</li></ul>

Holding regular lectures, presentations and open office hours in AI-related topics for startups, scaleups and other companies associated with Hetch Community, Hetch Incubator and Hetch Accelerator.

<h3>Solution &amp; objective</h3><p>Planned and executed the ”Explore Data Architecture” project by utilizing the knowledge, experience and expertise inherent in NordAxon as one of the main AI/ML professional services providers on the market.</p><p>&nbsp;</p><h3>Value</h3><p>After finalizing the project students learned:​</p><ul><li>Of work processes and differences between automation, personification, machine learning and AI along with an overview of how to solve existing market problems and areas of application in various industries applying AI.​</li><li>To identify AI business cases and to plan, identify resources, and perform specialized tasks needed to lead a project for the development of a product, service, or process which involves AI.​</li><li>The foundations of data architecture and how it’s composed of models, policies, rules or standards that govern which data is collected, and how it is stored, arranged, integrated, and put to use in systems and in organizations.​</li><li>To perform research and analysis work to collect data, process it and determine its quality of the data.</li></ul>

Provided industry leader advisory and tutorship for 2022 graduate students of Hyper Island. Expertise was given in the field of Data Management and Governance for Applied AI, through advising engagement in the ”Explore Data Architecture” project being one of the building blocks of educational program ”AI Business Consultants”.​

<h2>Challenges</h2><ul><li>Detecting fraudulent documents.</li><li>Working with very sensitive data, necessitating high security and access control during development and deployment of the solution.</li><li>Exploring trustworthiness of an automatic fraud detection tool.</li></ul><h2>Solution</h2><ul><li>A proof of concept tool proving that a number of key fraud scenarios are able to be automatically detected.</li><li>Using exploratory data analysis to extract information from documents and find signals indicating fraud.</li><li>A human-centered system focusing on interpretable computer vision by presenting automatically generated analyses together with the information and signals they are based on to the end users.</li></ul><h2>Value</h2><ul><li>Proven that it's possible to automatically detect key cases of fraud for this use case.</li><li>Decreased the need for manual human inspection and analysis.</li><li>Delivered a white-box computer vision system ready for live testing and evaluation of the effectiveness and interpretability of the system.</li></ul>

Using a data-centric approach and close collaboration with the customer we developed a computer vision system for automatically detecting fraudulent documents.

<h2>Challenges</h2><ul><li>Taking part in a highly exploratory project within a large business and quickly grasping their business needs.</li><li>Exploring the possibilities and limitations of currently available data.</li><li>Analyzing vast amounts of abstract product data, while ensuring the interpretability of the analysis.</li></ul><h2>Solution</h2><ul><li>An extensive exploratory data analysis to investigate the data, including creating interpretable data visualizations.</li><li>Rapid iteration and prototyping to explore the feasibility of several possible solutions, both AI-based and using classical statistics.</li><li>Close collaboration with the client team, assuring that the results are well communicated and presented.</li></ul><h2>Value</h2><ul><li>A thorough investigation the possibilities and limitations of the currently available data.</li><li>A resource for exploring business cases of other available data in the department.</li></ul>

Using exploratory data analysis, it was investigated whether data science-based techniques could be used to predict future device failures based on available testing data upstream.

<h3>Challenges</h3><ul><li>That the images available are taken at angles that complicate the out-of-the-box use of pre-trained object classifying models​.</li><li>That the license plate character recognition is complicated by the fact that the images have a wide range of brightness over the sets. ​</li></ul><p>&nbsp;</p><h3>Solution</h3><ul><li>Fine-tuning object recognizing models YOLO and Bounding-Boxes by training it on a dataset with images off vehicles passing the bridge.​</li><li>Adapting thresholding and edge recognition to each specific image, and by that ensuring that noise is filtered out, but the important information is kept.</li></ul><p>&nbsp;</p><h3>Value</h3><ul><li>Lowering error-rate of vehicle classification impacting maintenance statistics as well as customer billing.​</li><li>By creating a model with increased image analysis capabilities and benchmarking against the current system, we give the customer a possibility to evaluate their current solution.</li></ul><p>&nbsp;</p>

In this project, we are focusing on image analysis of the vehicles passing the Øresund bridge. By combining several image pre-processing techniques and pre-trained image analysis models, we are aiming to present a final model that can analyse the available data even further than what is being done now.

<h3>Solution &amp; Objective</h3><ul><li>AI is an important component in building the banks of the future. Unlocking the value from AI will be the differentiator between the winning and losing banks. One of the major blockers for achieving this is among employees very limited understanding and knowledge about AI as technology.​</li><li>The bank's country manager announced in Dagens Industri (2019) her ambition to AI Educate/ Upgrade all employees in the bank by offering them basic AI education about.​</li><li>The bank ordered creation of 1h webinar, a concise and basic version of AI Fundamentals customized towards banking industry. “AI för All” webinar demystifies AI and in brief explains most important aspects and building blocks concerning AI technology.​</li></ul><p>&nbsp;</p><h3>Value</h3><ul><li>After taking the webinar employees gained beyond the hype understanding about AI and ML, allowing them to when so required in their role with confidence engage in AI related discussions and decisions.​</li><li>Employees were introduced to:​<ul><li>Brief history, present and future of AI​</li><li>Insights about how machines are trained to make predictions and importance of data in this process​</li><li>Most common applied AI/ML use cases within Finance, Banking, Insurance, Manufacturing &amp; Supply chain, Automobile, Healthcare, Retail, Agriculture and Public Sector.​</li></ul></li></ul>

Designed, created material and recorded customized and white-labeled AI introducing webinar for the bank's approximately 1.000 employees; “AI för Alla” fundamental AI training.​

<h3>Challenges</h3><ul><li>By attending classical SFI courses new residents in Sweden take long time to learn Swedish language sufficient for finding a work in Sweden and properly integrating in Swedish society.​</li><li>In AI solutions, using rule-based engines for language learning is limiting natural conversations among its users.​</li><li>In the “jungle” of Conversational AI providers on the market, identifying a tool with intelligent enough engine for teaching its users learning a language.​</li></ul><p>&nbsp;</p><h3>Solution</h3><ul><li>Set up robust agile way of working with all participants involved in the project and technical platform provider leading the project implementation towards a solution properly tested by its future users, SFI students.</li></ul><p>&nbsp;</p><h3>Value</h3><ul><li>Improved language learning among SFI students through larger exposure to spoken language.​</li><li>Faster integration in the Swedish society, increased chances of getting employed and all the other socio-economical benefits associated with learning the native language in the new country of residence.</li></ul>

In this explorative project we analysed market looking for AI-driven tools used for improving learning Swedish language. Provider with a rule-based solution was selected and we lead technical PoC implementation of this language engine solution intended for usage by the SFI (Swedish for Foreigners) students.​

<h3>Challenges</h3><ul><li>Time and stakeholder management needed for getting necessary security and risk approvals to upload and explore the bank's masked internal data in the cloud.​</li><li>Some resistance and skepticism from Quantitative Research and Analyses team members towards GCP as a platform.​</li></ul><p>&nbsp;</p><h3>Solution</h3><ul><li>GCP provides number of cloud services including computing, data storage, data analytics and machine learning. Platforms capabilities exploration was performed with quantitative analyzing uses cases based on internal Fixed Income transactions data in combination with external Refinitiv (Bloomberg) market data.​</li><li>Final delivery included GCP instructions webinar and balanced scorecard comparing banks current analytical, infrastructure and on-prem storage capabilities with corresponding ones in GCP.</li></ul><p>&nbsp;</p><h3>Value</h3><ul><li>The bank's investments in physical hardware (storage, CPUs etc.) and software are high. PoC demonstrated option for cost savings with GCP which offers fast speed, scalability (unlimited storage), reliability and mobility for affordable pricing plan; users pay monthly as per their usage. E.g. option to per demand combine banks internal data with external heavy market data is a great cost saving for the bank, or any other bank.​</li><li>Access to latest releases of analytical applications, tools and frameworks on the market, in one place.</li></ul>

Responsible for coordination and execution of PoC initiative with objective to explore and compare GCP (Google Cloud Platform) capabilities with the bank's Analytical, Storage and Infrastructure tools used by the Big Data Trading Analytics, Quantitative Research department. ​

<h2>Challenges</h2><p>Manual and time-consuming efforts are required by companies to perform sustainability verification EU-CSRD (Corporate Sustainability Reporting Directive). Various documents must be read and textual responses formulated as&nbsp;a&nbsp;part of the verification. Using excelfiles as a verification tool is not optimal but need for an intelligent portal is clear.</p><h2>Solution</h2><p>The portal's front and backend are developed in React and Python and solution is deployed in GCP (Google Cloud). The user interface, UX/UI design, was carried out in close collaboration with the customer, with mock-ups created with Figma tool. AI-generated textual responses to the verification questions functionality is based on prompting the large-scale language model (LLM) &nbsp;with verification-relevant documents uploaded and indexed in a vector database and LangChain framework.&nbsp;</p><h2>Value</h2><p>Well designed portal with supporting AI-functionality reducing manual and time-consuming efforts&nbsp;during&nbsp;sustainability verification process.&nbsp;Easier and faster verification means that more companies are verified.</p>

Development of AI-driven portal for sustainability verification of companies.

<figure class="table"><table><tbody><tr><td><p><strong>Challenge</strong><br>Identifying water quality in numerous private wells around Sweden is not always applicable and open data for all&nbsp;those is not available. &nbsp;There is hypotheses about whether available open data can be used to accurately interpolate the geospatial distribution of chemical properties across different homogeneous groundwater areas in Sweden.<br>&nbsp;</p><p><strong>Solution</strong><br>A pipeline workflow was setup for consuming, cleaning, analysing and visualizing available open data aiming to investigate which properties are suitable to interpolate and what timespans are suitable. Analysis conducted was based on previous scientific research, with interactive visualization tools. Final kriging workflow was setup to project the interpolated area on top of an interactive map for granular analysis.<br>&nbsp;</p><p><strong>Value</strong><br>VANA's information portal aims to protect values and promote balanced interactions among Earth's species. It allows environmental experts to share critical knowledge and information, some of which may be endangered, and earn by providing insights needed for environmental permits. Vana´s clients benefit from quick assessments of site suitability, like environmental impact statements. Adding to Vana´s portal geospatial Interpolation of chemical properties in groundwater based on open data&nbsp;opens for evaluation of this feasibility study and creates great value for clients that need this data. Necessity for more dense data became apparent during the feasibility study, which could be arranged by increasing number of users along with suggestions on possible next step focusing on how to improve the data workflow.</p></td></tr></tbody></table></figure>

Feasability study on the application of geospatial interpolation using kriging techniques on chemical properties of groundwater based on open and semi-open data. 

<h2>Challenges</h2><p>Existing speech recognition solutions offered as&nbsp;cloud-based&nbsp;APIs from the tech giants such as Google, Microsoft and AWS are less good at understanding/transcribing low-resource languages like Swedish and even worse at pronunciations and dialects.&nbsp;Transcribing conversations that may contain sensitive and personal data, e.g. healthcare visits,&nbsp;is usually not allowed in&nbsp;the cloud but only on local on-prem servers. Using tech giants APIs for&nbsp;speech recognition is also a continuous subscription cost.</p><h2>Solution</h2><p>About 2.5 hours of voice data was collected from people with Swedish as a second language via a website created by NordAxon for this purpose. The web was explored for suitable open datasets with varying Swedish dialects and two were chosen (Common Voice and NST) for building the prototype.</p><p>Research was conducted&nbsp;in the field of open-source STT (Speech-to-Text)&nbsp;models and best performing&nbsp;for the project were selected (ContextNet &amp; VoxRex) and fine-tuned with collected data. Results from&nbsp;our&nbsp;models were evaluated with Word/Character/Meaning Error Rates (WER, CER, MER) and compared to the results from the leading cloud-based&nbsp;API&nbsp;solutions on the market (GCP, Azure, AWS).One of prototypes demonstrated that&nbsp;finetuning an open STT model with only 2.5 own data&nbsp;outperforms Google GCP and&nbsp;is just as good as Microsoft Azure.&nbsp;Conclusion is made that with&nbsp;around additional 100h of&nbsp;"Swedish as a second language"&nbsp;data&nbsp;Azure API can be outperformed as well.</p><h2>Value</h2><p>Speech recognition model that also understands&nbsp;pronunciations and dialects is of great value for many digital applications used by Swedish municipalities. On-prem STT models can also be used on sensitive and personal data, being GDPR-compliant, and subscription cost associated with cloud-based solutions is eliminated.</p>

Implementation of on-prem Speech-to-Text model finetuned on crowd-collected "Swedish as a second language" voice data.

<h2>Challenges</h2><p>Companies manufacturing packages,&nbsp;such as FrontPac, are depending on&nbsp;reoccurring processes for reading&nbsp;information&nbsp;about chemicals, allergens etc. from various Safety Data Sheets, which require manual and monotonous efforts by employees.&nbsp;</p><h2>Solution</h2><p>The solution is using&nbsp;PDFMiner&nbsp;for reading texts from Safety Data Sheets&nbsp;PDFs&nbsp;arriving to one of the FrontPac mailbox.&nbsp;RegEx is used for extracting relevant information&nbsp;from this raw text&nbsp;before automatically being sent and&nbsp;filled in documents relevant for the production process.</p><h2>Value</h2><p>Replacing&nbsp;manual&nbsp;flow&nbsp;with&nbsp;automated reduces the need for human intervention. Accuracy of readings improves as operational errors become fewer; automated processes do not have "tired eyes".</p>

Short summary: Implementation of solution that automatically reads PDFs arriving in a mailbox.

<h2>Challenges</h2><p>Older industrial machines used for rubber production often do not have open interfaces for reading and extracting machine settings and continuous values being measured. Instead, the machines screens displaying the settings and measurement values must be visited by humans physically and values are noted manually. This setup is both time-consuming and results in frequent faulty noted measurement data.</p><h2>Solution</h2><p>A web-camera is set up for continual filming the machine screen displaying the machine settings and various measurement values. The prototype solution is using computer vision and OCR for processing live frames from the camera video feed while saving the resulting OCR data,&nbsp;time-series&nbsp;data of the machine’s settings and its measured values, to an SQLite database. That data is thereafter used for various production analyzing purposes. Solution is dockerized and enabled for live-testing via a Nvidia Jetson Orin industrial computer and using an external IP-camera as camera device.</p><h2>Value</h2><p>Having an automated setup for monitoring and analyzing machine settings and measurement values is saving human time and enabling data analyses, maintenance and diagnostics for the machine maintenance and production flow purposes.</p>

Prototype automatically OCR extracting/reading data from an industrial machine monitor via a video feed from a web-camera.

<h2>Challenges</h2><p>Soeco Kontorsmöbler is a local reseller of used office furniture. Manually identifying the various office furniture in connection with inventory is a time-consuming task.</p><h2>Solution</h2><p>A large dump of various office furniture images was used as a reference database. Natural Language Processing techniques “Embeddings and Similarity Search” were applied in the implementation of the AI classifiers as a part of wider Inventory App development.</p><h2>Value</h2><p>More efficient inventories which is saving time/money and giving employees an opportunity to spend time on other more exciting tasks.</p>

Implementation of AI-solution for classification of office furniture during inventory workload.

<h3>Solution &amp; Objective</h3><ul><li>The main objective with our classroom format AI/ML Fundamentals course is to remove the hype and the black box around AI and ML (Machine Learning).</li><li>One course session is 6-7h long with a mix of theory and hands-on training in Google Colab. Participants get to build and trains their own ML models, no programming skills as prerequisite.​</li><li>Course material is comprehensive but well-structured and provided by highly pedagogical and practically experienced tutors.​</li></ul><p>&nbsp;</p><h3>Value</h3><p>After finalizing the course, attendees revealed that they:​</p><ul><li>Were able to understand AI and ML beyond the hype and felt much more confident and prepared for the AI related decisions and actions on their ongoing AI/ML journeys, with business benefits as outcomes.​</li><li>Could better understand data scientists and data engineers in their teams/projects and expected more efficient collaboration with them.​</li><li>Were aware of the hidden costs in the AI projects, understood the importance of having good quality data and gained proper knowledge about the AI/ML workflow and what it takes to train and build decision forecasting machines, accurately.​</li></ul>

We provided one day AI/ML Fundamentals training course for the company's Supply Chain & Production C-level Board of Leaders and Directors. Session was held in Paris, during teams one-weeks executive team building days.​

<h2>Challenges</h2><ul><li>Distilling the essence and most important practical knowledge to teach students about large and complex fields, while still being digestible and pedagogical for students.</li><li>Joining as external domain experts into a running educational program, and adapting the technical level of the workshops to the students' knowledge levels.</li></ul><h2>Solution</h2><ul><li>Two full day digital workshops. One day focusing on Computer Vision and one day on Natural Language Processing.</li><li>Combining lectures on underlying theory with interactive practice through hands-on coding labs in jupyter notebooks.</li><li>Raise a conversation and discussion on ethical and moral risks of ML model behaviors.</li></ul><h2>Value</h2><ul><li>A concise overview of two very important fields in AI, a short history as well as current opportunities and limitations.</li><li>Giving students an interactive walk-through of a practical workflow: preprocessing, model development and model evaluation.</li><li>Discussing the ethics, risks and possibilities with AI/ML technology.</li></ul>

As supplier of external domain experts in AI and ML, we complemented an educational program in programming for AI with workshops in Computer Vision and Natural Language Processing.

<h3>Challenges</h3><ul><li>Time and stakeholder management needed for getting approvals and extracting data from various data sources in the bank.</li><li>Quality of and accessibility to datasets consumed by the monitoring solution.</li><li>Complex technical architecture required by the solution, number of new tools and platforms introduced.</li></ul><p>&nbsp;</p><h3>Solution</h3><ul><li>Contextual Decision Intelligence platform Quantexa was implemented, used for detecting and identifying indicators of anomalous and suspect activity by applying holistic data using network analytical techniques and behavioral and peer group methods.</li><li>Project management was executed in a highly agile way, with 2-weeks sprints and three steps approach: Pilot, MVP and the third phase for further onboarding of other asset classes.</li></ul><p>&nbsp;</p><h3>Value</h3><ul><li>The bank's money laundering scandal arose in 2017-2018 around €200 billion of suspicious transactions that flowed through the Estonia-based bank branch of the bank from 2007 to 2015.</li><li>Fines of several billion dollars have been payed to financial regulators in Denmark, US and other European countries and upon new breaches more fees will be required.</li></ul>

Carried responsibility for technical project management in the banks AML (Anti Money Laundering) Trade Flow Monitoring project with rule-based technique used for detecting and alerting financial crime suspected trades.

<h2>Challenges</h2><ul><li>Planning and creating course content for a new program in practical data science. Making sure that the contents of each course are relevant for the industry needs. This was the first time this program was held in Skåne.</li><li>Getting students ready for practical machine learning and data science from scratch in a short period of time.</li></ul><h2>Solution</h2><ul><li>Planning and teaching 4 courses during the students first year of the two-year long program.</li><li>Building the education gradually. Starting with the basics of Python for machine learning, project management and the theory behind machine learning, and culminated in a self-led project course where the students got to apply everything they had learnt in a practical setting.</li></ul><h2>Value</h2><ul><li>After our block of the education, the students had gone from little-to-no programming and machine learning experience to completing their own data science projects.</li><li>The students gave us an average evaluation score of 8 / 10 over the four courses.</li></ul>

Planning, teaching and creating course content for several courses for a Higher Vocational Education in data science. The courses were focused on programming in Python for machine learning, project management for data science-projects and both theoretical and applied machine learning.

<h3>Challenges</h3><p>To speak in Swedish is&nbsp;a&nbsp;challenge for&nbsp;immigrants in Sweden. Typical dialogues in society, such as job interviews, care visits, conversations with educators at (pre)school are extra stressful situations for this target group.&nbsp;There is also higher unemployment&nbsp;associated&nbsp;within this target group, due to the lack of experience and fear of speaking in Swedish.</p><h3>Solution</h3><p>Emely´s language generator is based on a OpenAI GPT Transformer LLM &nbsp;prompted for various dialogue situations.&nbsp;Multiple other&nbsp;AI APIs are integrated into the solution; Google Translate API for translations&nbsp;and&nbsp;Azure Speech-to-Text API for Emely's speech synthesis. The front and backend are developed in React and Python and the solution is&nbsp;deployed in Google Cloud. Emely is NordAxon's own service product. Operation, maintenance, and new releases are handled by the NordAxon team.</p><p>For more information check Emelys infosite www.emely.se.</p><h3>Value</h3><p>Emelys&nbsp;NLP-driven language generator&nbsp;encourages conversations with the users by continuously asking questions suited to the ongoing type of dialogue, for example questions typical during a job interview. In this way, users are without fear for wrong-sayings and any other pressure trained to speak&nbsp;in Swedish and get better at it over time.</p>

AI-driven language learning assistant Emely offers situational dialogue training for example job interviews, care visits, easy-Swedish dialogues and various other conversations that arise in society. 

<h3>Solution &amp; objective</h3><ul><li>AI is an important component in building the organizations of the future. One of the major blockers for achieving this is among employees very limited understanding and knowledge about AI as technology.​</li><li>The institution ordered creation of 90min broadcasted AI Fundamentals Intro session for all their resources engaged in IT. Content was customized so that it answers taxation AI/ML concerns and use cases . Main objective was to demystify and in brief explain most important aspects and building blocks concerning AI technology.​</li></ul><p>&nbsp;</p><h3>Value</h3><ul><li>After attending the live session employees gained beyond the hype understanding about AI and ML, allowing them to when so required in their role with confidence engage in AI related discussions and decisions.​</li><li>Employees were introduced to:​<ul><li>Brief history, present and future of AI​</li><li>Insights about how machines are trained to make predictions and importance of data in this process​</li><li>Most common applied AI/ML use cases in their or related industries​</li><li>Advices about how to go about when planning and executing AI/ML projects​</li></ul></li></ul>

Designed, created material and broadcasted via Skype customized and white-labeled AI/ML Fundamentals Intro training (90min) for the institution's 250 employees.​

<h3>Solution &amp; Objective</h3><ul><li>AI is an important component in building the organizations of the future. Unlocking the value from AI will be the differentiator between the winning and losing companies. One of the major blockers for achieving this is among employees very limited understanding and knowledge about AI as technology.</li><li>Kraftringen ordered a creation of 1-2h webinar, a concise and basic version of AI Fundamentals customized towards the energy sector. The “AI för All” webinar demystifies AI and in brief explains most important aspects and building blocks concerning AI technology.</li></ul><p>&nbsp;</p><h3>Value</h3><ul><li>After taking the webinar employees gained beyond the hype understanding about AI and ML, allowing them to, when so required, engage with confidence in AI related discussions and decisions.​</li><li>Employees were introduced to:​<ul><li>Brief history, present and future of AI​</li><li>Insights about how machines are trained to make predictions and importance of data in this process​</li><li>Most common applied AI/ML use cases within Finance, Banking, Insurance, Manufacturing &amp; Supply chain, Automobile, Healthcare, Retail, Agriculture and Public Sector.​</li></ul></li></ul><p>&nbsp;</p><p><br>&nbsp;</p>

Designed, created material and recorded customized and white-labeled AI introducing webinar for Kraftringen.​

<h3>Solution &amp; Objective</h3><ul><li>The main objective with our classroom format AI/ML Fundamentals course is to remove the hype and the black box around AI and ML (Machine Learning).</li><li>One course session is 6-7h long with a mix of theory and hands-on training in Google Colab. Participants get to build and trains their own ML models, no programming skills as prerequisite.​</li></ul><p>&nbsp;</p><h3>Value</h3><p>After finalizing the course, attendees revealed that they:​</p><ul><li>Were able to understand AI and ML beyond the hype and felt much more confident and prepared for the AI related decisions and actions on their ongoing AI/ML journeys, with business benefits as outcomes.​</li><li>Could better understand data scientists and data engineers in their teams/projects and expected more efficient collaboration with them.​</li><li>Were aware of the hidden costs in the AI projects ,understood the importance of having good quality data and gained proper knowledge about the AI/ML workflow and what it takes to train and build decision forecasting machines, accurately.​</li></ul>

During three separate full day course sessions we provided AI/ML Fundamentals training course for approximately 100 leaders and decision makers, among those Group CIO and other C-level leaders board members.​

<h3>Challenges</h3><ul><li>No open nor known to exist, labelled dataset for Swedish ASLR. Great volumes of data are needed for creating an accurate model, Swedish sign language counts approximately 120.000 signs.</li><li>Limited insights about sign language grammar and how this language is used in practice e.g. speaking with two vs one hand, dialects etc.​</li><li>For the next phases, more sophisticated PoC extension and MVP, assuring required funding.</li></ul><p>&nbsp;</p><h3>Solution</h3><ul><li>PoC indicated that it is possible to build a reliable ASLR model with most promising ML architecture solution consisting of two steps, model for human pose detection and sign recognition from skeletal 2D-/3D-points.</li><li>PoC data consisted of very limited number of signs taken from two sources: 1) Sign Language Linguistic University in Stockholm and 2) Video recorded by NordAxon data scientists and sign language translators, in collaboration.​</li></ul><p>&nbsp;</p><h3>Value</h3><ul><li>According to SCB(Statistics Sweden) ≈ 18,5 % of Swedish grown-up population is diagnosed with some kind of loss of hearing. Providing this big minority group automated 24/7 available sign language translating solution will have great impact on their quality and safety of lifes.​</li><li>Great cost reduction for the society, sign language translators cost 800 kr/h exkl. VAT.​</li></ul>

Investigated feasibility for building an ASLR (Automatic Sign Language Recognition) AI/ML model for Swedish sign language. ​

<h3>Challenges</h3><ul><li>Due to limited knowledge majority of Swedish leaders and decision makers perceive AI/ML as a hyped black-box technology. That is one of the main blockers preventing Sweden from becoming one of the globally leading AI players.</li><li>Removing silos, replacing competitive attitude with collaborative and creating unified “AI language” for academia, public and private sectors - on the country level.</li></ul><p>&nbsp;</p><h3>Solution</h3><ul><li>AI Innovation of Sweden is a national and neutral initiative, functioning as an engine in the Swedish AI ecosystem. The focus is on accelerating the implementation of AI through sharing knowledge and data, co-location of competences, and collaboration projects, all with a strong emphasis on ethics, transparency, and security.</li><li>90 people from industry, academia, science parks, municipalities and regional organizations based in south Sweden attended the pre-study workshops and aligned on how south node should be organized.</li></ul><p>&nbsp;</p><h3>Value</h3><ul><li>According to Gartner, AI-derived business value is forecasted to reach $3.9 trillion in 2022 with three different sources of value: customer experience, new revenue, and cost reduction.</li><li>Sweden is a small country but with great AI potentials, given historical and current technological progress. For achieving this aligning all main AI players, on the country and regional level, is crucial.</li></ul><figure class="image"><img className='content-img'  src="https://firebasestorage.googleapis.com/v0/b/nordaxon-webapp.appspot.com/o/AI%20Se%20-%20Case.jpg?alt=media&amp;token=6889fd5a-5601-4d4e-a24a-3a97e2668395"></figure>

Provided advisory on opportunities with AI/ML and maturity needed for that, from technical and implementational practice aspects. Assisted organizing and conducting workshops with south of Sweden key AI stakeholders. Emphasized crucial need for Swedish decision makers to have Data and AI strategies in place and their skillset AI-upgraded if Sweden is to be considered as one of the globally leading AI players.

AI/ML Case Studies — Client Work from the Nordics