Data Science UA

Tasks

• Develop machine learning models to analyze client behavior, segment users, and predict preferences.
• Build a personalized recommendation engine for workout plans, services, and offers.
• Create a scheduling optimization model to align client availability with trainer resources.
• Integrate a large language model as the conversational core of client’s in-app virtual assistant.
• Fine-tune the LLM to reflect the company’s brand tone, gym-specific procedures, and customer support guidelines.
• Build an ETL pipeline to unify data from client’s mobile app, gym systems, and CRM/ERP into a centralized repository.
• Implement secure cloud storage for models, logs, and backups using AWS S3.

Challenges

• Historical records were incomplete or noisy, limiting early model accuracy.
• Client preferences shifted over time, requiring continuous model updates, while new clients or services lacked sufficient data for meaningful recommendations.
• The LLM occasionally generated plausible but incorrect outputs.
• The platform needed to comply with data privacy regulations such as GDPR while maintaining user trust.

Solutions

• Client data used for modeling was anonymized using salted one-way hashes; personally identifiable information was excluded from analytics workflows.
• Custom instructions, prompts, and reinforcement feedback were used to shape the assistant’s responses, ensuring brand consistency and accuracy in gym-related queries.
• Built a robust data pipeline to extract, clean, and centralize information from multiple systems, improving data reliability for ML models.
• Combined behavior prediction, segmentation, and recommendations into one agent that could adapt based on real-time user feedback and platform usage patterns.
• The mobile app was updated with a clear privacy policy and tools for users to view, export, or delete their personal data.

Outcomes

• Successfully deployed a secure and compliant AI assistant that recommends content, manages schedules, and engages clients in personalized conversations.
• Improved user satisfaction through more relevant recommendations and faster response times.
• Maintained data privacy and built trust with clients through transparent handling of sensitive information.
• Laid a scalable foundation for ongoing machine learning improvements and LLM updates.

Tasks

• Build a machine learning model to forecast price movements across selected trading instruments.
• Develop an AI agent that interprets model outputs to inform automated decision-making processes.
• Integrate reporting functionality to generate and deliver performance metrics and KPIs.
• Implement an external trigger system (e.g., email and webhook) to automatically activate the agent and handle follow-up responses.

Challenges

• Financial markets are unpredictable; historical patterns may not hold during major geopolitical events or policy changes, especially within the UK market.
• Extracting predictive signals from structured market data and unstructured sources like news or social media required significant effort, with no guarantee of impact.
• Embedding robust trading safeguards (e.g., stop-losses, position limits) that adapt dynamically to model confidence and market volatility.
• Models trained on historical data could fail in live trading environments, particularly when trained on limited market cycles.

Solutions

• Developed adaptive models with online learning capabilities to regularly retrain on new data and adjust to changing market trends.
• Integrated macroeconomic indicators and real-time event data into the model to improve responsiveness to market shifts.
• Used automated feature engineering (AutoFE) tools to rapidly test and refine a wide range of input variables.
• Worked closely with financial experts to vet model features and interpret market behavior accurately.
• Deployed advanced NLP pipelines to extract signals from news headlines, social sentiment, and regulatory updates.
• Hosted models on high-performance infrastructure, optimizing serving latency and throughput using FastAPI.

Outcomes

• Delivered an AI agent capable of handling real-time price predictions for multiple instruments.
• Enabled automatic activation via email or webhook, with the agent retrieving inputs, executing analysis, and sending back KPI reports.
• Reduced manual intervention in forecasting and reporting workflows.
• Created a scalable foundation for automated decision-support tools in financial operations.

Tasks

• Develop advanced computer vision models to detect protocol violations in aseptic environments.
• Build an AI agent to analyze visual input in real time, interpret events, and issue alerts based on risk levels.
• Implement a centralized dashboard for event logging, reporting, and trend analysis.
• Enable multi-channel notifications with escalation logic for unacknowledged alerts.
• Ensure the system meets pharmaceutical industry compliance and integrates with the company’s existing infrastructure.

Challenges

• The client operates under strict regulatory requirements and high safety standards. Maintaining false positive and false negative rates under 8.5% and 9.8%, respectively, was essential to avoid alert fatigue or missed violations.
• Accurate labeling of complex actions across large video datasets was time-consuming but critical to model performance.
• Cleanroom environments introduced challenges like reflections, inconsistent lighting, and occlusions, making reliable detection more difficult.
• Continuous monitoring across multiple camera feeds generated significant data volumes that needed to be stored, retrieved, and audited efficiently.
• Manual oversight was not scalable, and traditional systems lacked the responsiveness needed to actively reduce incidents or generate usable compliance data.

Solutions

• The team annotated over 7,000 cleanroom images and used targeted data augmentation and synthetic generation to simulate real-world conditions: glare, shadows, occlusions.
• Image processing pipelines were tuned to normalize lighting fluctuations before model inference, improving detection stability.
• The AI agent used a configurable rules engine to evaluate combinations of detections and escalate based on predefined risk scenarios (e.g., lack of goggles near sterile zones).
• Developed an automated logging mechanism that records all safety-related events, enabling detailed audit trails for regulatory reporting.
• Each alert included timestamped, role-specific information and direct references to SOPs, helping staff act quickly and appropriately.

Outcomes

• Deployed a real-time sterility monitoring system across the client’s manufacturing zones.
• Enabled fast, accurate detection and escalation of non-compliant procedures, reducing human oversight burden.
• Maintained target false alert rates, balancing safety and efficiency.
• Provided traceable audit logs and incident reports to support internal QA and external regulatory inspections.
• Built a scalable foundation for continuous model improvement and future expansion to additional compliance areas.

Tasks

• Improve access to learning materials for platform users.
• Convert unstructured content (PDFs and video presentations) into a searchable, structured format.
• Enable fast and intuitive information retrieval through natural language queries.
• Lay the groundwork for scalable AI-based learning tools.

Challenges

• Much of learning content presented was locked in static formats like PDFs and recorded videos.
• The lack of content structure made it difficult to implement intelligent search or any personalized assistance.

Solutions

• Used OCR to digitize and structure course content from PDFs and video transcripts based on predefined business logic.
• Implemented a vector database to index content semantically.
• Deployed a Retrieval-Augmented Generation (RAG) system using the o3-mini large language model to enable contextual responses to user queries.
• Built a simple, secure interface with Streamlit for learner access.
• Hosted all infrastructure in AWS Cloud for scalability and performance.
• Designed the solution as an MVP, with ongoing iterations based on user feedback.

Outcomes

• Reduced time learners spend searching for relevant materials by over 70%.
• Made course materials accessible via conversational queries, improving user interaction and satisfaction.
• Established a solid infrastructure foundation for future AI features.
• Continued platform development is supported by real-world usage insights and feedback from learners.

Tasks

• Establish a single client-consultant touchpoint for client questionnaires.
• Enable advanced analytics for consultants.
• Generate HRPP, Salary reports, and Graded Pay Structure.

Challenges

The main challenge was to create a system that ensures continuous availability, enhances customer experience, boosts consultant effectiveness, and minimizes their data-related tasks.

Solutions

• Developed a secure, single-touchpoint service with dynamic dashboards and drill-down capabilities.
• Built an autonomous system with a modern, user-friendly interface.
• Used historical data for predictive analytics and personalized recommendations.

Outcomes

Improved customer satisfaction with a user-friendly service, enhanced consultants’ effectiveness, enabling greater project scalability, reduced consultants’ data tasks, freeing resources, and showcased EY’s commitment to modern solutions and brand value.

Tasks

• Build the first core engineering team of AI/ML specialists.
• Identify top-tier candidates with expertise in cutting-edge AI research and development.
• Establish long-term collaboration on hiring for R&D teams to support innovation.

Challenges

• Defining highly specific criteria for candidates aligned with Reface’s unique product development goals.
• Attracting top talent in a competitive AI landscape.

Solutions

• Conducted a detailed analysis of Reface’s technical needs and long-term R&D objectives.
• Implemented a targeted recruitment strategy leveraging industry networks and AI-specific talent pools.
• Created a streamlined hiring process to evaluate both technical skills and cultural fit.

Outcomes

• Successfully hired Reface’s first core team of AI/ML engineers.
• Established a long-term partnership for ongoing AI talent acquisition.
• Supported Reface in maintaining its position as a globally recognized leader in AI-driven personalized content creation.

Tasks

• Develop AI algorithms with dashboards and web portal development
• Implement and optimize CV algorithms for complex industrial environment cases
• Data preparation and annotation for industrial use cases.

Challenges

The project involved developing multiple algorithms and solutions. Additionally, implementing and integrating web portals and dashboards with the platform required extensive data preparation and annotation.

Solutions

• Designed and optimized CV algorithms for complex industrial scenarios
• Integrated real-time location systems (RTLS) and other sensors with CV outputs
• Developed dashboards and web portals for seamless usability.

Outcomes

• Achieved design, implementation, and optimization of CV algorithms for complex environments
• Successfully developed and optimized pipelines for performance improvement
• Delivered minimal involvement in day-to-day tasks, thanks to local leadership and efficient solutions.

Technologies used

• Python, TensorFlow, PyTorch, C++
• RTLS, Lidar, 1080p camera, thermal camera

Tasks

• Develop a predictive model on the wind energy production forecast
• Analyze historical data for accuracy
• Implement and test the model

Challenges

Accurately forecast wind energy production based on historical data.

Solutions

• Developed a predictive model for wind energy production
• Utilized historical data for training and testing the model
• Implemented machine learning techniques to enhance forecast accuracy

Outcomes

• Highly preciseodel forecasts – with Mean Absolute Percentage Error less than 10%.

Client:
Ukrainian startup that provides market research services in the retail industry. The company collects data from all the popular retail stores in Ukraine and sells insights extracted. In particular, the company provides price-setting assistance based on competitor analysis

Challenge:

The company has encountered a problem while adding new products to its database. A simple comparison did not work since identical products may have different names in different stores, and not all stores provide complete information about the product's characteristics. Therefore, a more profound approach was required. 
 

Solution:
Our team has developed an algorithm based on word embeddings (vector representations of words). Using BERT-based library sentence-transformers, we incorporated the power of natural language processing for this task. First, by calculating the embedding of the product name and then comparing it with the embeddings of products in the database, we managed to find items with a certain similarity threshold.

Results:
The developed algorithm can now quickly find similar products in a vast database, and it is possible to set a threshold for the desired similarity. The system also extracts all the possible characteristics from the product name (such as size, brand, and country of origin).
 

Technology stack:
Python (nltk, sentence-transformers, scipy, pymongo), MongoDB, Docker.

Tasks

• Enable real-time audio and text recognition
• Provide instant insights for sellers
• Develop a new feature for customizing video and audio content

Challenges

Enabling real-time audio and text recognition, providing instant insights for sellers. Developing a new feature for customizing video and audio content, aiming to enhance the platform’s functionality.

Solutions

• Released real-time audio and text recognition & processing to offer immediate information to sellers
• Developed a new feature of video and audio customization
• Data annotation for multilanguage audio data

Outcomes

• Immediate information delivery through real-time recognition
• Enhanced platform functionality with customizable video and audio content
• Multilanguage data annotation for broader reach

Client

A national store network with hundreds of locations in Ukraine. The field is highly competitive with several other similar networks.

Challenge

Production sale prices are changing with time. Sometimes they are lagging behind the demand and competitors. The network wanted the exactly optimal prices to maximize the revenue, which was challenging to do manually.

Solution

The customer turned to Data Science UA to develop such a system. The design took into account multiple surrounding factors. It distributed the goods in categories and adjusted prices for each of them. The result fitted the price-demand distribution to improve the revenue for each product.

Technology stack

Python

Client:

A big taxi management company with reviews in Ukrainian, English, etc.
 

Challenge:

Our main challenge was to build a reliable classification system that would allow the internal analyst team to quickly and efficiently analyze text reviews in multiple languages. Additional challenges were: text quality (grammar, spelling issues distort prediction quality), multilingual pipeline (no out-of-the-bag model can equally well work with different languages), quality control (business-specific metrics had to be defined to assess the performance), and custom-tailored categories (not just general ones, but also those defined by industry).
 

Solution:

First, we built a preprocessing pipeline with language detection, spelling correction, machine translation, etc. Then, on top of it, a custom model based on TCN layers was trained to detect business-specific categories in the reviews. Translation quality, precision, and recall metrics for each specific category were checked with the predefined business goal.
 

Results:

A fast, reliable tool that allowed the client’s quality assurance team to quickly, easily analyze driver’s performance, provide overall user satisfaction metrics, promptly solve critical issues.
 

Technology stack

Python, NLTK, Transformers, Tensorflow, etc.

The client's global mobile banking app puts bill payment, banking, credit, investments all in one single platform. It connects to users' existing bank accounts, allows the user to pay all bills from one place. The main task for Data Science UA was to create and use an ML model for the classification of all transactions. So you as a user have an access to the necessary tools to stay on track with your financial goals in a cost-efficient way. Users can pay their bills and track important financial milestones through one integrated experience.​

Client:

Eastern review aggregator company that assembles text reviews from multiple resources (Google reviews, Yelp, etc.).

Challenge:

Build and deploy a model to categorize English text reviews according to the business-predefined categories. The main challenge was to build a relational category scheme with multiple levels (first being the most abstract and all consequent ones more discrete) and then train a custom model to predict categories in the text according to the scheme adequately. Additionally, a small model to highlight category-specific text in the review had to be implemented.

Solution:

First, we built a preprocessing pipeline with grammar correction, tokenization (with menu items treated as a special case), etc. After that, training data describing the complex category scheme was collected and properly labeled. On it, a custom BERT-based model was trained. Using text-distance measures simple text highlighting model was built, in order to immediately highlight predicted categories in the text.

Results:

After deployment, the client successfully integrated the model into their service as a dashboard, so that restaurant chains could easily filter, compare and analyze the performance of individual restaurants. The highlighting model was also used to provide accessible feedback from the review text directly

Technology stack:

Python, NLTK, SpaCy, Transformers, Tensorflow, scheduling, MySQL.

Tasks

• Audit current IT architecture.
• Analyze reporting processes across departments.
• Recommend solutions for analytic function optimization.
• Gather requirements for reporting and analytics.
• Define master data systems for each category.
• Propose new IT architecture for system integration.
• Outline development plan for the new system.

Challenges

The main challenge was to develop a comprehensive strategy to automate data collection and reporting processes, integrate various data systems, and improve data-driven decision-making capabilities for Servier Ukraine.

Solutions

• Established unified reporting standards for all data categories.
• Created a single Master Data system per category.
• Enabled automated data exchange across systems to reduce manual tasks.
• Built an automated reporting system for recurring reports.
• Recommended MS Power BI as the primary reporting tool.

Outcomes

• Automated data collection and reporting.
• Enhanced data integration and minimized silos.
• Strengthened decision-making with predictive analytics.
• Reduced manual tasks, boosting data efficiency.
• Improved user experience with a modern interface.

Our client is the largest national retail chain of beauty and health stores, offering more than 30,000 assortment items. It has more than 1000 stores operating throughout Ukraine. The client’s goal was to up-sell and cross-sell goods through a recommendation system and to make the realization of clients’ churn prediction modeling.

During the collaboration with Data Science UA the client achieved x2 improvement of key metric - Baseline model (recommending only the most popular SKUs) had MAP@10 = 0.03. This means that out of 10 suggested SKUs client is likely to buy 3. Our model had MAP@10 = 0.07. This means that out of 10 suggested SKUs client is likely to buy 7.

Architecture that was used – LSTM (neural network).