Николай Деревяга

AI/LLM and localization engineer with 20 years in digital and media products. At TONLY.AI, I designed and run a production AI localization platform for 23 languages, with five pipelines covering content from UI strings and articles to books and video. The platform processed more than 17 million characters in 2025, and its text localization track cut unit costs by 87%.

I built a multi-agent LLM workflow that routes work through Analyzer, Translator, Editor, and QA agents, with automated quality gates and human review where it matters. Two pipelines run fully autonomously, while a quality evaluation layer uses automated scoring, LLM-as-a-judge, structured JSON reports, and a three-axis rubric calibrated with 30+ professional translators.

For video, I automated the path from speech recognition and timed subtitles through translation and voiceover to assembled video. I also developed an internal React service for managers to run that workflow, and maintain the production system single-handedly, including its containerized services, scheduling, retries, and spend limits.

Beyond localization, I build LLM applications and agent infrastructure: RAG assistants with embedding-based vector search (including an employee onboarding assistant), MCP integrations and my own ecosystem of Claude skills with role-based agents, and generative video pipelines (Runway, Kling, VEO, Wan Animate 2 in ComfyUI, Remotion, ffmpeg) with self-hosted inference on GPU servers.

Before TONLY.AI, I led the volumetric video studio at Spherum and worked on VR and AR products, including seven apps shipped to the App Store and Google Play. I co-founded AVATunes, which reached #2 in the Russian App Store with about 400,000 downloads, and previously ran a digital agency with international projects. My earlier work also included sound engineering in radio and television.

Based in Moscow; open to remote work and relocation (work visa sponsorship required).