AI systems can scrub sensitive data from operational files before records circulate externally. The Institute of Smart Systems and Artificial Intelligence at Nazarbayev University (ISSAI) unveiled Qupia to automate this workflow. The solution was showcased at AI & Digital Bridge 2026 in Astana.
According to ISSAI, Qupia automatically flags personally identifiable information (PII) and facilitates pseudonymization. The engine processes text streams, structured documents, static images, and audio files.
Replacing Names and National IDs with Tokens or Redactions
Within the Qupia interface, users can select alternative data treatments. Across text documents, identified fields can be masked or substituted with placeholder tokens like “Name” or “IIN”. For audio files, sensitive voice segments are replaced with acoustic bleeps or silent frames.
The core objective of automated redaction is preserving the analytical context of an artifact while stripping personal identities. For instance, customer support teams can analyze feedback patterns without storing customer identities. This serves as an illustrative deployment scenario rather than formal audited performance metrics.
Currently Operating in Demo Mode
The platform maintains an important operational constraint: ISSAI explicitly categorizes Qupia as a technology demonstrator. Algorithmic parsing may occasionally miss sensitive data fields, leading developers to advise against uploading actual production credentials.
Users are instructed to conduct manual human-in-the-loop validation prior to disseminating processed records. The release of Qupia therefore does not ensure complete automatic anonymization of any uploaded payload.
The official ISSAI release lacks precision and recall metrics or third-party audit results. Without statistical validation, assessing how robustly the model handles varied document layouts and noisy recordings remains difficult.
ISSAI Showcased Broader Model Portfolio
During the summit, the institute demonstrated MangiSoz (speech and text translation spanning 11 languages), TilSync (real-time captioning and translation), the multimodal foundation model Oylan, and ISSAI Voice Studio for vocal synthesis.
Furthermore, ISSAI finalized cooperation memorandums with the National Bank of Kazakhstan’s Digital Development Center, QAZBOT TECHNOLOGIES, and China’s iFLYTEK. These alliances support ISSAI’s broader research footprint; they do not indicate active enterprise integration of Qupia across these partners.
RAEM.KZ previously profiled the research team behind these models in our feature on ISSAI director Atakan Varol. The next milestone for Qupia will be releasing verified accuracy benchmarks and commercial licensing terms beyond its initial demo interface.
