Alona Liuzniak

Dentists spend most of their time focusing on what matters most—treating patients. As a result, detailed documentation often has to be completed under time pressure, even though it is essential for accurate records and proper billing.

To support this process, I developed an intelligent system that reviews dental treatment documentation and compares it with best-practice treatment patterns for specific procedures. This analysis identifies potential gaps or inconsistencies in the documentation and provides structured suggestions for improvement. To ensure patient privacy, I implemented an automated anonymization step using a small language model (SLM) that removes sensitive data before the records are passed to the main AI models for analysis. In addition to developing the solution itself, I coordinated development activities to ensure the system could be implemented efficiently and aligned with the needs of the medical professionals who would use it.

The system helps dentists create clearer and more complete records, improving documentation quality and ensuring that performed treatments are properly reflected in billing while reducing time spent on manual checks.

Many design teams rely on time-consuming manual checks to ensure their work adheres to internal UX guidelines. I developed an AI-powered platform that automatically evaluates designs against these standards directly during the design process.

The solution combines an automated evaluation engine with a Figma plugin and a web interface. Designers can run guideline-based tests directly on their designs, allowing the system to identify inconsistencies, missing elements, or potential usability issues early in the process. This reduces the need for repeated manual reviews and helps teams maintain consistent design standards across products.

To further support designers, I built an interactive AI assistant that can answer questions about both the current design file and the organization's UX guidelines. By using a knowledge-based AI approach, the assistant provides contextual guidance and helps designers quickly understand how to apply internal standards in practice.

The result is a tool that integrates directly into designers' workflows, helping teams work faster, maintain high design quality, and scale UX governance across the organization.

In this project, I led the development of an intelligent chatbot solution tailored for customer support and product recommendations in the chemistry industry. The chatbot leveraged state-of-the-art RAG to process vast amounts of product information sourced from the company's website and internal documents. Within just 3 months, I successfully delivered a fully functional, end-to-end solution that smoothly integrated with the company's workflows.

By employing advanced techniques such as prompt engineering, text summarization, and a 2-layer search system, I significantly improved the chatbot's ability to deliver precise and contextually relevant responses. The innovative search mechanism combined product name filtering with relation matching, enabling the chatbot to handle complex multi-hop questions effectively. To ensure transparency and user trust, I incorporated RAI mechanisms, providing clear justifications for product recommendations and aligning the solution with the company's ethical standards.

In today's fast-paced business environments, communication workflows often rely on multiple data formats, including text and images. To address this challenge, I enhanced a chatbot solution with an image-ingestion pipeline, enabling it to analyze and integrate visual data into its knowledge base. This breakthrough enabled users to ask questions not only about textual content but also about images, making multimodal question answering a reality.

Beyond the chatbot's ability to process and retrieve visual data, I also developed Outlook and Microsoft Teams connectors, enabling flawless integration with enterprise communication workflows and making the chatbot more customizable and accessible to users by allowing them to interact with it directly within familiar tools. This integration streamlined workflows, reduced friction in information retrieval, and enabled organizations to tailor the chatbot's responses and capabilities to their specific needs.

To ensure transparency, I applied Explainable AI (XAI) techniques, providing users with visibility into how image processing workflows and retrieval mechanisms operate. This ensured that every chatbot response based on visual data was traceable and verifiable, reinforcing reliability.

In the world of financial services, ensuring compliance while providing a seamless user experience is a delicate balance. The Know Your Customer (KYC) process, a critical requirement for bank account openings, often suffers from manual inefficiencies and fraud risks. My goal was to automate and optimize KYC verification by leveraging deep learning and computer vision.

I led the development of an AI-driven KYC solution that streamlined identity verification, enabling customers to onboard faster, more safely, and more smoothly. A major milestone was improving the recognition accuracy of German ID security features by 60%, thanks to advanced deep learning models and computer vision techniques.

This project highlighted my ability to bridge AI innovation with real-world business challenges, applying expertise in ML, neural networks, and advanced data analysis to build a highly effective, fraud-resistant KYC automation system. By combining technical excellence with a deep understanding of financial security needs, I helped financial institutions enhance trust, security, and efficiency in their customer verification processes.

Fraud detection in banking is a high-stakes challenge, where every false positive disrupts legitimate customers, and every false negative costs millions. My mission was to enhance fraud prevention while simplifying complexity, turning an unmanageable 10-million model setup into a highly efficient 9-model solution.

The breakthrough came from an optimized customer segmentation strategy, which allowed us to dramatically reduce the number of required models without compromising predictive accuracy. This streamlined model training, deployment, and evaluation makes fraud prevention scalable and cost-effective.

Beyond segmentation, I enhanced the existing XGBoost model performance by engineering new, highly predictive features, improving fraud detection accuracy while minimizing unnecessary transaction rejections. To further elevate anomaly detection, I implemented an autoencoder-based strategy for transaction evaluation, introducing an additional layer of unsupervised fraud detection that identified subtle, previously undetectable fraud patterns.

In modern robotics, efficiency and adaptability are key to automating complex tasks, especially in manufacturing and logistics. My mission was to create intelligent robotic systems that could perceive, learn, and act autonomously, leveraging reinforcement learning, synthetic data, and real-time inference.

To train AI models for robotic object picking, I designed and generated synthetic training data by creating artificial 3D environments that simulated real-world tasks, eliminating the need for expensive real-world data collection and making the system scalable and highly adaptable.

The real challenge, however, lay in optimizing motion planning to ensure the robot could pick objects quickly, accurately, and in dynamic environments. By applying reinforcement learning, I trained the system to adapt its movements in response to feedback, thereby enhancing both efficiency and decision-making.

This project combined my expertise in computer vision, reinforcement learning, and ML, demonstrating how AI can enhance real-world automation. By integrating synthetic data and adaptive learning, I built a solution that made robotics smarter, faster, and more efficient, setting the stage for the next generation of intelligent automation.