Blog

What we ship, and why.

Product updates, engineering notes, and opinions from the team building the pipeline. No growth-hacking listicles.

47 posts·23 product update·20 engineering·4 perspective

Product UpdateOctober 9, 2026· 4 min read

Product Update — Every language in one launch, comments on the page, and PDFs quoted before you take them

One upload can now go into several languages at once, you can look at the translated page next to the original and comment on it, and a PDF can be measured before anyone commits to a price. Plus two fixes we owe a write-up.

Read

Product UpdateOctober 7, 2026

Comment on the page, not the segment

Reviewers see segments. Clients see pages. Compare layout puts the original and the translated page side by side, and lets you draw a box on the page, say what is wrong, and get a rebuilt file back.

Read

Product UpdateOctober 6, 2026

One upload, every language

A brochure going into Spanish, French and Chinese used to be three uploads, three quotes and three emails. Now it is one launch: one job per language, each with its own assets, one quote, one email, one download.

Read

EngineeringOctober 6, 2026

Eleven pipelines that only looked new

An InDesign file took 22 minutes. It should have taken half that. Eleven file formats had the current pipeline’s name and the old pipeline’s engine, and every Chinese, Japanese and Korean job was quietly doing its work twice.

Read

Product UpdateOctober 5, 2026

Quote the PDF before you take the job

A partner asked for something simple: tell me what a PDF will cost before I accept it. That turned into three things: PDF Quote, an OCR service that keeps the page, and translations that come back in the original layout.

Read

EngineeringSeptember 29, 2026

“Chinese (Simplified)” is not “Chinese”

Some Simplified Chinese jobs came back mostly in Traditional characters, with only the glossary terms in Simplified. The cause was one line that removed the part of a language name in parentheses, and a tiebreak that depended on which server ran the job.

Read

Product UpdateSeptember 18, 2026

Product Update — Fuzzy matches, scoped glossaries, and images straight into OCR

Fuzzy matches are now edited instead of shipped untouched or thrown away, glossary terms know where they apply, loose images go straight to OCR, and reviewing a ZIP ends in a ZIP. Most of it started as a client report.

Read

EngineeringSeptember 18, 2026

The ZIP that finalized everything except itself

Every file in the batch was correct. The batch was not: Finalize on a reviewed ZIP never finished, and every ZIP project reviewed since July had the same ending. Then a second bug at the same seam, from the other direction.

Read

EngineeringSeptember 12, 2026

A glossary term has a scope

An entry that says “health → salud” is right until it meets a product name. How we taught glossary enforcement where a term applies: protected names win, the longest match wins, and proper names are left alone.

Read

Product UpdateSeptember 11, 2026

A fuzzy match is a draft, not an answer

A 92% match is an approved translation of a sentence that is almost the one in front of you. We were either shipping it untouched or throwing it away. Now it is edited: keep the memory’s wording, change only what the source changed.

Read

EngineeringSeptember 9, 2026

The tags were right. Our cleanup moved them.

A Phrase Word file came back with nearly every inline tag piled up at the end of its segment. The model had placed them correctly. A defensive cleanup deleted them, and a repair step put them back where it could.

Read

EngineeringSeptember 2, 2026

Column[1]Count: when a tag is a line break

In a software UI file, a tag glued between two words was a line break inside a label, and we were translating the label half by half. The fix was easy. Switching it on everywhere was the mistake worth writing about.

Read

EngineeringAugust 20, 2026

The setting that never reached the engine

You asked us to leave your 100% and 101% matches alone. The job recorded the choice and the engine never saw it, so 230 approved segments were rewritten. Fixing that uncovered a second bug, and the second one lived in six pipelines.

Read

Product UpdateAugust 19, 2026

Support: an answer in minutes, not on Thursday

Every support ticket is now acknowledged within minutes, naming the job and what happens next — and the person who picks it up already has a diagnosis attached. Median first response went from about sixty-three hours to under twenty minutes.

Read

Product UpdateAugust 17, 2026

Subtitles and closed captions, from the video itself

Send the video, not an SRT somebody else made. Cues are built from word-level timing, translated inside their own time budget, reviewed against the picture, and delivered as SRT, WebVTT and TTML — with optional burn-in.

Read

EngineeringAugust 17, 2026

We read the whole style guide. All 144 rules.

“Netflix-compliant” usually means 42 characters and two lines. We inventoried the five published documents, checked our pipeline against every rule in them, and found the ones we were quietly failing.

Read

Product UpdateAugust 16, 2026

A subtitle editor with the video in it

Reviewing subtitles in a text editor is guessing. The review screen now has the player, a waveform with the cues on it, shot marks, keyboard retiming, back-translation per cue — and a rebuild that bakes exactly what you saw.

Read

EngineeringAugust 12, 2026

The Finalize button that finalized nothing

Our audio review editor let you correct a transcript, rename the speakers, and click Finalize. The deliverable was regenerated from none of it. Here is the fix, and the word-level playback that came with it.

Read

Product UpdateAugust 11, 2026

The portal now takes the order — end to end

A week ago the client portal was where your clients received work. Now it is where they start it: requests with deadlines and reference files, a queue on your side, live status on theirs, and delivery file by file.

Read

Product UpdateAugust 11, 2026

One review, five reviewers, zero collisions

Big jobs used to have an obvious bottleneck: review was one person's job. A review can now be split across reviewers by segment ranges — with the server, not politeness, keeping them out of each other's work.

Read

Product UpdateAugust 11, 2026

Your client sees your price. Never our credits.

Most of our customers resell translation. The portal now prices work in your rates and your client's currency, and produces monthly statements from what was actually charged — your margin stays your business.

Read

EngineeringAugust 11, 2026

The report card was buried in the ZIP

Every delivery ships with machine QA reports — glossary compliance, memory usage, tag integrity. They lived inside the delivery package, which is a polite way of saying nobody read them. Now they are on the screen. Surfacing them also meant auditing them.

Read

EngineeringAugust 10, 2026

We audited our own translation memory. It needed it.

A translation memory is a promise: a sentence approved once comes back the same way forever. We traced every path in ours to check the promise was being kept. Some of what we found was not flattering. All of it is fixed — and now you can verify it yourself.

Read

Product UpdateAugust 5, 2026

One door, not fifteen connectors

Our competitors count connectors. We shipped a documented REST API with signed webhooks instead — because the system you actually use is never the sixteenth one on their list.

Read

EngineeringAugust 5, 2026

Fifty files, one upload, one delivery

Batch processing sounds trivial until a single file in the middle fails, or fifty files come back translating the same sentence four different ways. Here is what a batch actually has to survive.

Read

Product UpdateAugust 4, 2026

A portal for your clients, under your brand — and still under your control

Your clients stop emailing you files. You keep the approval, the price ceiling and the brand. Every order arrives as a request you accept or decline — with a guaranteed maximum charge attached.

Read

PerspectiveAugust 4, 2026

A pass rate is not a quality score

Everyone reports a QA pass rate. Almost nobody can tell you how bad the failures were. We shipped MQM-style scoring with weighted severities — and one distinction the rest of the industry skips.

Read

Product UpdateAugust 4, 2026

Paying reviewers for the work the AI didn't do

When AI does the first pass, per-word rates stop making sense. New: effort per reviewer and per job, with typing a correction counted differently from clicking accept — plus CSV export and a printable report with the evidence.

Read

Product UpdateAugust 3, 2026

Product Update — Enterprise sign-in, an API, and a portal for your clients

Single sign-on with Microsoft Entra ID and Okta, two-factor authentication your organization can require, a public API with signed webhooks, and a branded portal where your own clients submit work.

Read

PerspectiveAugust 3, 2026

Your 100% matches might be your biggest liability

The industry bills 100% TM matches at 0% and passes them through untouched. That's a great deal — unless the TM is wrong.

Read

EngineeringAugust 3, 2026

Tracked changes that survive translation

A client edits the English doc with Word tracked changes. You already translated v1. They want v2 in Spanish — with the same tracked changes applied. No tool did this. So we built it.

Read

Product UpdateAugust 1, 2026

Product Update — July 2026

Self-serve billing, a reviewer role for your client's own editors, Human QA on a toggle, OCR that keeps the images, and portal-only delivery for compliance.

Read

EngineeringJuly 22, 2026

Human QA: what happens when the machine isn't enough

Back-translation you can hand to compliance, and a professional linguist one toggle away. How the human layer works — and why our reviewers cannot close a job without proving they read it.

Read

Product UpdateJuly 15, 2026

Already translated? Then we only fix the errors

Send a bilingual file somebody else translated. We correct accuracy, terminology, grammar and numbers — and leave the style alone, so your linguist's voice survives. Billed at a reduced rate.

Read

PerspectiveJuly 8, 2026

Three branches, fifteen formats: what you can actually upload

CAT files, Office documents, native and scanned PDFs, images, audio, subtitles. A map of what Fily accepts, what comes back, and why text, scans and audio each need their own engineering.

Read

Product UpdateJune 30, 2026

Product Update — June 2026

A model refresh across every pipeline, the Golden TM that builds itself, Human QA as an orderable service, and a glossary generator that starts from your archive.

Read

EngineeringJune 24, 2026

Spanish is not one language, and neither is yours

es-MX, es-ES and es-US are not interchangeable, and treating them as one locale is how a health plan ends up addressing its members with the wrong word. How locale rules actually work here.

Read

Product UpdateJune 20, 2026

You already have a translation memory. It's just scattered.

Two tools that turn old deliveries into working assets: an aligner that builds a TMX from a source file and its translation, and a generator that extracts a glossary from documents you already sent.

Read

Product UpdateJune 10, 2026

Review without a CAT tool

Invisible doesn't mean opaque. A real word-level diff, cross-file consistency across a 50-file batch, back-translation for languages you don't read — and the data that comes out the other side.

Read

Product UpdateMay 30, 2026

Product Update — May 2026

Every pipeline rebuilt on the v2 core, a 3-layer glossary autofix with rollback, client rules that cascade five levels deep, and the tracked-changes pipeline nobody else had.

Read

EngineeringMay 12, 2026

Reading what nobody else can read: OCR for faxes, scans and phone photos

Generic OCR fails silently on a 150-DPI medical fax — it returns a blank page and no error. Two independent vision engines, page-level retry, layout reconstruction, and images kept in place.

Read

Product UpdateApril 30, 2026

Product Update — Spring 2026

One pipeline architecture for fifteen formats: the spring rebuild that turned a collection of tools into a system.

Read

EngineeringApril 15, 2026

A glossary that fixes, not flags

Every tool detects terminology violations and hands you a list. Ours rewrites the target — in three layers, each one validating tags and rolling itself back rather than breaking your file.

Read

EngineeringMarch 20, 2026

Audio is not a document: why transcription needs its own pipeline

Running audio through a document pipeline produces transcripts nobody can file. Speakers, verbatim fidelity, timestamps, and branded templates are the actual product.

Read

EngineeringMarch 5, 2026

What actually happens to your file: the twelve steps

Upload, download, done — but not a black box. A walkthrough of all twelve pipeline steps, in the order your file meets them, and what each one is protecting you from.

Read

EngineeringFebruary 18, 2026

Inside the pipeline: how a Trados file survives AI translation

Inline tags are where AI translation quietly ruins CAT files. How Fily parses, protects, repairs, and re-verifies them — so Trados accepts every file we return.

Read

PerspectiveFebruary 2, 2026

Why we built Fily

Every tool in this category sells you a workflow to operate. We wanted the translation. So we built the pipeline we could not buy — and this is what we believe about how translation should work.

Read