Roman Vainshtein

Planning failures in multilingual systems can be systematically diagnosed and mitigated, leading to significant performance gains in low-resource languages.

LLM vulnerability scanners can give you wildly different security scores depending on *who* is judging the attacks, not just the model's weaknesses.

Stateful defenses are not just better, but *necessary* to protect multimodal RAG agents from sophisticated attacks that bypass stateless filters.