Enhancing AI-Driven User Interactions through Data Segmentation in the MyInkarne PoC Project
Choudhury Zahid · Doria (University of Helsinki) · 2026
The rapid evolution of AI-driven applications has underscored the critical role of data management in balancing personalization, performance, and privacy. Systems like MyInkarne, which enable users to create AI-powered "digital twins" for posthumous interactions, face unique challenges: processing heterogeneous data types (text, voice, images), ensuring GDPR compliance, and maintaining low-latency responses. While feeding raw, unstructured data to external AI models like OpenAI could improve response accuracy, this approach risks privacy violations and inefficiency. This thesis investigates how structured data segmentation, partitioning data into contextually relevant subsets, can optimize AI-driven interactions while adhering to ethical and regulatory standards. The study evaluates the impact of segmentation on AI response accuracy, latency, and user trust in the MyInkarne system. Key methodologies include role-based segmentation (e.g., isolating "Creator" training data from "Visitor" interactions), topic-based memory categorization (e.g., "family" or "career" tags), and hybrid architectures combining localized data indexing with cloud-based AI synthesis. Experimental results demonstrate that structured segmentation reduces OpenAI API calls by approximately 40%, improves retrieval efficiency by around 25%, and enhances perceived authenticity through more targeted memory retrieval and grounded response generation while persona conditioning and voice-related presentation further supported perceived authenticity. By addressing gaps in real-time processing, standardization, and privacypersonalization trade-offs, this research establishes best practices for data segmentation in AI systems. The findings contribute to GDPR-compliant frameworks for digital legacy applications, offering scalable solutions for healthcare, e-commerce, and other domains reliant on secure, personalized AI interactions.