Dmitrii G.

I am a localization engineer. For the last five and a half years I have worked on domain machine translation for real client content: memories, termbases, tags, and engines that have to stay consistent. At Custom.MT I trained and scored production engines for life science, industrial, and UI accounts, and wired them into Trados, Phrase, memoQ, and custom APIs so project managers could run MT without leaving the tool. The source data was whatever the client sent: noisy translation memories, broken tags, short strings. Quality was a fixed held-out set, SacreBLEU and COMET, not a sample someone happened to post-edit. LexaForte (lexaforte.com) is that same pipeline on my own stack. I clean the client TMX, fine-tune a small open-source NMT model, convert it to CTranslate2 so it serves on CPU, and lock terminology and tags before and after the pass. An LLM post-edit is optional, and it checks rather than rewrites. The highest score so far is 73.5 SacreBLEU on 2,000 held-out pharmacology segments, EN–RU, against DeepL Pro at 32.0 and Google AutoML at 56.5 on the same sentences. Across 13 other real tests the hybrid beat the other providers. Training runs on AWS, GCP, or Azure.