Notes from the workshop
Things we learn while building Transept and watching how people translate.
Latest
No italics, no spaces, no あなた: what Japanese taught us about localization
We localized Transept into Japanese in a day and thought ten earlier languages had prepared us. They hadn’t. Fake italics, a space two style standards disagree on, words that mean something else in Japanese, a katakana name with two spellings, and why our assistant must never sound like an anime heroine.
More writing
Can AI detectors spot an AI translation? We tested Pangram and GPTZero
We translated passages of a novel between Ukrainian, English and Chinese with AI and ran them through Pangram and GPTZero. Every sample came back human-written, a plain ChatGPT translation included. Here is why, and why you should still disclose an AI translation.
Machine Translation Post-Editing (MTPE): What It Is, How It Works, and How to Do It Better
A practical guide to machine translation post-editing (MTPE): light vs. full post-editing, why raw MT still needs a human, how quality is measured — and why the usual batch-then-scrub model is often the wrong shape, plus the iterative, memory-fueled workflow that beats it on both quality and cost.
Ten languages in three days: field notes from localizing an AI translation product
We localized Transept into German, Ukrainian, Chinese, Portuguese, French, Spanish, Czech, Italian, Polish, and Turkish in one long weekend, using machine translation the way we tell our users to use it: with context, terminology decisions, and human post-editing. These are the plural rules, register flips, typography inversions, and hreflang lessons we collected on the way.
Literess as an agent: the editor who remembers your decisions and does the work
Most AI in translation tools is a chatbot bolted on the side. We built Literess as an agent instead — grounded in the same decision-context memory the product runs on, and able to take real actions on your behalf, always with your confirmation.
Translation Memory: What It Is, How It Works, and Why It Matters for AI Localization
A translation memory stores the lines your team already approved. We needed to figure out what TM means when the translator is an LLM, how the tools on the market remember (or forget), and the bet we made at Transept: memory as decision context.
How much does AI translation actually think?
We tried to control Gemini 3's thinking budget the same way we did with Gemini 2.5. It doesn't work that way anymore. Here's what we measured instead — and what it means for AI translation quality and cost.
Welcome to The Journal
Why we started writing here, what you can expect to find, and a small invitation to come back.
The writers
New entries land roughly monthly.
RSS