Admir Krnic

Neural audio watermarking research. Part-time, held concurrently with the role below. • Formalized backward compatibility in zero-bit neural audio watermarking — to our knowledge the first systematic study of the problem, and a requirement under EU AI Act Article 50 and its Code of Practice. Introduced an Angular Diversity Loss that restores true-negative rates to 99–100% against out-of-chain embedders with no measurable cost to detection accuracy or audio quality, and showed the same mechanism enables revocation of leaked embedders. Evaluated across 8 model versions spanning DenseNet, ViTMAE, ResNet and AST architectures. First author; under review at IEEE WIFS 2026. • Improved watermark robustness using contrastive learning, adding an InfoNCE objective and projection head to Timbre Watermarking training. Achieved 95–99% bitwise recovery accuracy across 25 distinct distortion and attack types while preserving imperceptibility. First author; presented at a Montenegro AI Association (MAIA) workshop. • Currently extending the backward-compatibility framework to multi-bit watermarking. • Stack: PyTorch, AWS.

Pharmaceutical market analytics delivered as a platform-as-a-service; ~27 people. Previously Data Scientist. • Own the regulatory and supply-chain data domains end to end — from acquisition through SQL and Python transformation to the outputs served to the product frontend — across a pipeline processing petabytes per sync and billions of rows. • Rebuilt the drug-data parsing module from scratch, replacing a brittle approach that failed unpredictably on real-world inputs. The current system reaches 70–100% extraction success across covered sources and removed a recurring class of silent parsing failures. • Expanded regulatory coverage by 10+ countries and raised data quality across many existing ones, working from unstructured, contradictory web sources with no shared schema or canonical identifiers. • Redesigned the finished-dosage-form marketing authorization dataset from a flat to a nested structure and built the processing pipeline around it, enabling correct mapping of manufacturers to packs and substances — approaching 100% mapping accuracy on priority markets. • Unified US supply-chain regulatory data scattered across 3 sources with no join keys, resolving hundreds of thousands of entities through iterative analysis and heuristics. Cited by the company owner as materially increasing confidence in the US dataset. • Combine SQL transformation pipelines with Python and LLM-based processing into single reconciled outputs, using Postgres, Trino, dbt, SQLMesh and Temporal. • Interviewed and helped select a working student who was hired on my recommendation and now contributes to a project I initiated and continue to own.

• Built a content management system for internet service providers that controlled which banners were served based on the network gateway a user connected through — frontend, backend services and database schema. • Developed frontend and backend services for a mail administration application, letting clients manage mail domains, users and permissions. • Stack: Java (Spring), PHP, Angular, SQL, Cassandra, REST APIs.