Skip to content

Olga Beregovaya

VP of AI at Smartling

  • 1 episode on Chain of Thought
  • 3 topics covered
  • 1 full transcript
  • Apr 2025 first on the show
Olga Beregovaya

At the time of the conversation, Olga Beregovaya was VP of AI at Smartling. On Chain of Thought she traced the shift from rule-based NLP to LLM-driven translation and the problems that remain, including the English-centric nature of AI, domain-specific errors, and under-resourced languages.

Cover art for Breaking the Language Barrier: Smartling's AI Translation Pipeline | Olga Beregovaya

Latest episode with Olga Beregovaya · EP 20

Breaking the Language Barrier: Smartling's AI Translation Pipeline | Olga Beregovaya

Olga Beregovaya, Smartling ·

Show notes & transcript → All listening options →

Episodes

What the conversations cover

These notes summarize the episodes featuring Olga Beregovaya, including other speakers in panel conversations. Follow a source link to read the discussion in context.

EP 20 Breaking the Language Barrier: Smartling's AI Translation Pipeline | Olga Beregovaya

  • Olga Beregovaya benchmarks human output at “a translator can deliver 2,000 words a day — that’s the metric,” and says a well-trained, well-prompted, fine-tuned model lifts that by “magnitude and order of magnitude.” She frames the payoff as a package: lower cost, higher productivity, and more predictable quality. Episode takeaways
  • Beregovaya notes most foundational models remain “English language and English phenomena centric,” and, as a separate effect, a carefully crafted English prompt “will render a better result in the target language” than a linguist’s prompt in their own native language. Episode takeaways
  • Stuffing every linguistic asset into a prompt is, in Beregovaya’s words, a “waste of fortune.” Rather than feed translation memories, glossaries, and 100-page style guides into a prompt, or build 30,000-token prompts, Smartling uses RAG to “only fetch what’s relevant” — “a huge, huge, huge” game-changer and “a great way of mitigating biases and hallucinations.” Episode takeaways

In their words

Purpose built single purpose model would provide much more predictable outputs and its performance and its behavior would be much more explainable because you build it for this particular task and for this particular domain.

Olga Beregovaya Breaking the Language Barrier: Smartling's AI Translation Pipeline | Olga Beregovaya · 28:34

Transcript excerpts; ellipses mark omitted words. Some automatically selected quotes have filler words removed.