AI translation pipeline · Zero workflow

The best translation workflow
is the one you don't see.

Drop the file. Get the file. Translation, glossary, TM, and 12 QA steps — without opening an editor.

0words translated0hours of audio0language pairs0%QA pass

No card. No demo form. No editor.

Live activity
Refreshes every 30s · last update 0s ago
Processing now
1jobs
Last hour
0words
Last hour
0jobs
Last 24h
5,144words

Real production data from the Fily engine. Updated every 30 seconds.

What you don't do with Fily

The whole category sells you a workflow to use. We take it out of your way.

  • No browser editor to open.
  • No segment-by-segment approval.
  • No pipeline to configure.
  • No glossary to copy by hand.
  • No broken tags to fix.
  • No mediocre MT to clean up.
  • No PO to send an agency.
  • No 5-day wait.

Upload. Download. Done.

Invisible doesn't mean opaque

What actually happens inside.

Three phases. Twelve quality steps. Every one logged, auditable, attached to your delivery as an HTML report.

Your file enters the pipeline
Phase 1Steps 1–3

Extract

Read the file. Decide what to translate. Pass through what shouldn't change.

  • 01Source parsing

    Parse XML / DOCX / MQXLIFF / PDF. Detect format, tags, locked TUs.

  • 02TM match pass-through

    100% and 101% matches bypass the AI byte-for-byte.

  • 03Segment translation

    Frontier LLM with prompt caching, per-segment context.

Phase 2Steps 4–9

Enforce

Apply rules. Repair tags. Replace forbidden terms. Catch what AI gets wrong.

  • 04Glossary enforcement

    Mutates target — preferred terms inserted, not just flagged.

  • 05Style guide application

    Voice, register, punctuation rules per client / project.

  • 06No-source-bleed check

    Detects untranslated chunks that slipped through verbatim.

  • 07Forbidden term replacement

    Brand-forbidden words rewritten to approved equivalents.

  • 08Tag count repair

    Inline <g>, <x>, {N> realigned. Trados/memoQ accepts on import.

  • 09Consistency check

    Same source span → same target across the document.

Phase 3Steps 10–12

Deliver

Normalize. Score. Reassemble. Ship the file plus the report.

  • 10Number & date normalization

    Decimal commas vs dots, date formats, currencies per locale.

  • 11Quality scoring

    0–100 score per segment + per pipeline step. Visible in QA report.

  • 12Bilingual reassembly

    Re-pack the file in its original format. Tags intact. Ready to deliver.

File delivered + QA report attached

Each step has a name, an owner, and a measurable check. The HTML report shows what fired, what passed, and what flagged.

One upload. Any format.

From sdlxliff to scanned PDF.

Tags intact. TM 100/101 untouched. Formatting preserved. The file you upload is the file you get back.

Three branches · One pipeline

Translation. OCR. Transcription. Each one engineered, not generalized.

Most AI tools do one of these. We built dedicated pipelines for all three — because text, scans, and audio fail in completely different ways.

Branch 1

Translation

Native pipelines per format. Tags preserved. TM 100/101 untouched. Glossary that mutates output, not just flags violations.

What makes it different

  • Dedicated pipeline per CAT format (SDLXLIFF, MQXLIFF, TXLF, RTF, XLIFF 1.2/2.0)
  • Tag count repair + xid normalization — Trados accepts every file we return
  • Hard TM 100/101 pass-through (never sent to AI)
  • 4-layer glossary cascade: system → org → client → sub-client → job
  • DOCX tracked changes round-trip across languages (industry first)
  • 12-step QA report on every delivery

Live numbers

31M+ words · 99 pairs · 93.2% QA

Branch 2

OCR

Read what nobody else can read. Phone photos, faxed medical records, blueprints, historical archives — all parsed cleanly.

What makes it different

  • Dual-backend vision-language OCR with automatic failover on errors
  • Page-level retry — one bad page doesn't kill the document
  • Block-level extraction with layout reconstruction (columns, tables, footnotes)
  • Handwriting detection + confidence flagging per region
  • Works on PDFs (scanned + native), single images, faxes down to ~150 DPI
  • Output as searchable PDF or DOCX with reconstructed layout

Live numbers

200+ scanned PDFs · 7.8M words extracted · ~31K pages

Branch 3

Transcription

Audio in, bilingual document out. Diarized. Timestamped. Ready to file with the reviewer or the court.

What makes it different

  • State-of-the-art ASR + speaker diarization (multi-speaker)
  • Verbatim mode for legal — preserves “um”, false starts, pauses
  • Per-speaker translation with surrounding context (not segment-by-segment)
  • Bring your own template — transcripts in your client's branded DOCX (logo, fields, footer)
  • Domain prompting (medical, legal, technical glossaries during ASR)
  • Word-level timestamps for subtitle workflows (SRT, CPS-aware)

Live numbers

530+ hours of audio · 2,393+ jobs processed

All three branches share the same 12-step QA pipeline. See all 15 formats →

What lands in your inbox

A file.
And a report you can hand to your client.

  • Auditable segment by segment
  • Glossary hits and misses, highlighted
  • Forbidden terms flagged with replacements
  • TM 100/101 segments marked as pass-through
  • Tag count diff per segment
  • Quality score per pipeline step

Average QA pass rate across 8,504 jobs: 93.2%

QA_report_job_3b9fa4e8.html
— QA Report — Job 3b9fa4e8 —
✓ Glossary
432 / 433 enforced
✓ Tags
1,586 verified
✓ Forbidden
8 replaced
⚠ Review
3 flagged
TM 100/101 pass-through
591 segments preserved
Overall: 94% pass · 1,586 segs · 18.5min

The diff nobody else offers

AI speed. Human accountability.

Every paid account has a dedicated Account Manager. Not a chatbot. Not a ticket queue. A human who knows your formats, your glossary, your deadlines — and has a direct line to engineering when something needs to change.

The speed of a tool

AI pipelines run in minutes. Doc3, Smartcat, DeepL get this part. Files in, files out — no human in the loop.

The accountability of a vendor

Agencies have AMs but the AM isthe bottleneck — every PO, every change goes through them. Fily AMs unblock, they don't gatekeep.

Direct line to engineering

When standard isn't enough, your AM brings engineering in. We've built custom pipelines for specific clients in days, not quarters. Not a sales conversation — actual code shipped.

Available on every paying tier. Not an enterprise-only upsell.

When standard isn't enough

Custom pipelines, built by engineers. Not configured by a sales rep.

The 15 format pipelines are the baseline. When a client's workflow doesn't fit, we build the pipeline. Here are three we actually shipped.

docx_tracked_delta_v2

Tracked changes that survive translation

An LSP came to us with this: "Client edits the EN doc with Word tracked changes, we already translated v1 to Spanish. They want v2 in Spanish with the same tracked changes applied — not a fresh re-translation." No tool on the market did this.

We built a delta-aware DOCX pipeline that ingests EN + redlines + ES_v1 and emits ES_v2 with the equivalent insertions and deletions in Spanish. Reviewer keeps the audit trail. Hours instead of days per round.

Industry first — read the DOCX page

pdf_qa_12step_v2 + image_qa

Dual-backend OCR for impossible scans

Healthcare LSPs were hitting walls with 150-DPI faxed medical records that generic OCR failed on silently — returning blank pages without warning.

We built a fallback chain: a frontier vision-language OCR as primary, automatic failover to an alternate backend on truncation or JSON errors. Page-level retry, confidence flagging, handwriting detection in the QA report.

See the scanned-PDF pipeline

4-layer cascade

Multi-tenant glossary cascade

An LSP with 40+ end-clients needed a glossary system where org-level terms, client-level overrides, and sub-client (brand) overrides all stack — without re-attaching glossaries to every job.

We built a 4-layer cascade: system → org → client → sub-client → job. Forbidden terms inherit and override. Style guides cascade. PMs configure once, every future job picks up the right rules automatically.

See the LSP use case

Bring your own template

Transcripts in your client's branded document

Sonix, Rev, Otter hand you their default DOCX. Then your PM spends an hour re-formatting into the client's template: logo, case number, signature block, footer.

Upload your client template once. Every future transcript for that client lands inside it — speaker labels, timestamps, translations placed exactly where the template expects them. Zero re-branding.

See the audio pipeline

Got an edge case Doc3, Smartcat, or your current vendor can't handle?

Talk to our Account Manager team

Built for the way you actually work

Whether you're a translator, a PM, or a buyer.

Translate. Don't post-edit.

Drop your sdlxliff, mqxliff, or txlf. Get it back with your glossary applied, tags intact, and TM matches untouched. You review the QA report, not 4,000 segments.

AI speed. Your judgment on what to keep.

Your PMs stop being dispatchers.

Bulk uploads, client-specific glossaries cascading from org → client → subclient, batch ZIP delivery, and a QA report your reviewer can stand behind.

Dedicated Account Manager. Real human. Not a chatbot.

No agency. No emails. No 5-day wait.

Upload directly. Receive the translated file plus an auditable report. HIPAA-ready. On-prem available. SLA on Enterprise.

Account Manager for every paid tier. Backed by humans.

Trusted by translators and the teams that hire them

I stopped opening Trados for first-pass translation. Fily returns the sdlxliff with my glossary already enforced. I review, I deliver.

Translator, healthcare specialist

We process 200K words/week through Fily before they hit our reviewers. Our PM team grew zero last year while volume doubled.

PM, mid-size LSP

For our 18-language community communications, Fily is the only tool that handles every format we need in one upload.

Director, K-12 multilingual programs

8,504 jobs processed · 99 language pairs · 530+ hours of audio

Pay-as-you-go or subscription · No seats · No licenses

Same engine for everyone. Just pick your scale.

Every tier gets all features unlocked — Translation, OCR, Transcription, glossary cascade, BYO templates, custom pipelines. The only thing that changes is volume and the kind of support you need.

Trial

See it work with one file

Free

no card · no commitment

  • All features unlocked
  • Up to 15,000 words
  • 12-step QA pipeline
  • QA report attached
Try free

Solo

For individual translators & small projects

Talk to us

pay-per-word or monthly subscription

  • All features unlocked
  • Translation · OCR · Transcription
  • All 15 file formats
  • 12-step QA report on every job
  • API access
Talk to us
Most LSPs land here

Team

For LSPs, agencies, in-house loc teams

Talk to us

pay-per-word or monthly subscription

  • All features unlocked
  • Higher monthly volume
  • Batch ZIP delivery
  • Glossary cascade (org → client → sub-client)
  • Bring-your-own audio templates
  • Dedicated Account Manager
Talk to us

Enterprise

For organizations with custom requirements

Custom

volume + compliance + custom pipelines

  • All features unlocked
  • Custom volume + dedicated capacity
  • On-prem deployment available
  • Custom pipelines built by engineering
  • SLA 99.9% · HIPAA / SOC 2 ready
Contact sales

Same pipelines. Same QA. Same formats. Talk to us and we'll match you with the right scale — pay-as-you-go if you want flexibility, monthly subscription if you want predictability.

Frequently asked

Upload. Download. Done.

Try Fily with one file. No card. No form. No editor.

15 formats · 99 language pairs · 93.2% QA pass · Auditable on every job