Nina Lopatina is a staff developer advocate at MongoDB focused on helping developers build AI agents that retrieve trustworthy information, act on enterprise data, and retain useful context. Her career spans neuroscience, language technology, machine-learning privacy, document processing, and persistent agent memory.
Lopatina began in neuroscience, studying reinforcement learning, decision-making, and how the brain represents tasks. She was first author of research on cognitive maps in the orbitofrontal cortex conducted through the National Institute on Drug Abuse and the University of Maryland School of Medicine.
Her subsequent machine-learning research addressed both model structure and model safety. At Lab41, she explored structured approaches to language modeling; she also coauthored studies of privacy attacks against speaker-identification systems and privacy risks associated with adversarially robust training.
At Spectrum Labs, she worked on language technology for identifying harmful online behavior, including challenges involving scarce training examples, class imbalance, and lower-resource languages. She later became head of innovation at Nurdle, focusing on specialized and synthetic data for production language models. At Unstructured, she worked in developer relations and connected document preprocessing with production retrieval applications.
Lopatina subsequently joined Contextual AI as lead developer advocate. At the 2025 AI Engineer World’s Fair, she and Rajiv Shah led a workshop on production-ready retrieval agents, demonstrating document ingestion, cross-document financial analysis, and evaluation grounded in actual source material. Her later work at MongoDB addresses how AI agents can incorporate operational context and retain information across interactions.
- Document-grounded AI agents: Lopatina uses financial reports and deliberately spurious correlations to test whether agents extract correct figures, distinguish correlation from causation, cite supporting material, and acknowledge when documents cannot answer a question.
- Natural-language unit tests: She evaluates generated answers against concrete criteria, including numerical accuracy, cross-document synthesis, evidentiary support, acknowledged uncertainty, and unnecessary verbosity. Those results can guide changes to prompts and agent configuration.
- Production-ready document retrieval: Her work emphasizes accurate extraction from complex PDFs, tables, and images alongside dependable source attribution and developer-accessible APIs.
- Persistent agent memory: At MongoDB, her focus extends retrieval into application architectures that give agents access to relevant operational data and useful context from earlier interactions.
