locAGI • Translation QA Workbench

From raw bilingual file to
client-ready delivery.

locAGI is the AI-augmented translation QA pipeline our linguists use every day. It turns any messy post-MT bilingual XLIFF — from memoQ, Trados, XTM, Phrase, or any CAT tool that exports standard XLIFF — into a scored, commented, client-ready deliverable. Humans stay in charge of every final call.

8
Pipeline stages
Chunked inference
Any AI model
MQM
Error framework
XLIFF
Any CAT tool
scroll ↓
01Termbases

Start with a single source of truth for terminology.

Before any AI touches a segment, we centralise every confirmed term, acronym, and forbidden phrase in a termbase. That way every downstream step — MT, AI post-editing, QA — speaks the same vocabulary as the client.

  • Import from TBX, XLSX or CSV in one click
  • Multi-TB merge at runtime — combine product + style TBs on the fly
  • Context & definitions per entry for disambiguation
  • Linked to templates so the AI always sees the right glossary
Open Termbases
🌐
Pharma_EN_DE.tbx
1,284 entries • 2 linked templates
ENDENote
contraindicationGegenanzeige§ regulatory
active ingredientWirkstoff
DO-NOT-TRANSLATE: MedPortalMedPortal🔒 brand
TBXXLSXCSVExport to CAT tool
02Templates

Bundle client rules into reusable AI templates.

A template is the client's style guide in one payload: free-text instructions, linked termbases, acronym handling, date / number / currency localisation, register, tone. Link a template once — every AI step after that behaves like the client expects.

  • Free-text instructions the AI reads verbatim each run
  • Linked termbases — pick one or merge several
  • Acronym & date rules applied before the AI replies
  • Client & domain tags so you find the right template fast
Open Templates
📋
ACME Pharma — DE formal
Client: ACME • Domain: Medical
// Instructions the AI sees:
Use formal German (Sie-form). Keep drug names in English
unless the TB says otherwise. Format all dates as DD.MM.YYYY.
Preserve inline tags {1}, {2}, ...
🌐 Pharma_EN_DE🌐 Brand_namesSie-formDD.MM.YYYY
03MT / AI — Band-aware

AI post-edits only what needs editing.

MT/AI reads the XLIFF TM match score for each segment and adapts intensity: high matches are barely touched, low matches get a full rewrite. 100% matches stay locked by default. You keep the leverage your linguists already earned in their CAT tool.

  • 100% locked — TM passthrough, never edited
  • 95–99% — micro-fix: casing, punctuation, tags
  • 85–94% — togglable band for client-specific review
  • <85% — full AI rewrite within template rules
Open MT / AI
🤖
Match-band policy
Applied per segment on import
🔒 100%
TM leverage
Locked — never edited.
⭐ 95–99
Micro-fix
Minor tag / spacing / case.
🟡 85–94
Togglable
Include for medical; skip for lit.
🔴 <85
Full rewrite
Template + TB applied.
04AIPE — Full AI Post-Editing

Heavy-lift AI post-editing with your model of choice.

When you want every segment rewritten to a target register — not just the low matches — AIPE runs the whole file through an AI model you pick, in chunked batches, building a per-file glossary as it goes so terminology stays consistent across the entire document.

  • Any provider — bring your own API key, swap per project
  • Private by default — run a local model on-prem when data can't leave
  • Per-file glossary rebuilt automatically each run
  • Smart retry on partial failures — no segment is dropped
  • Data-residency aware — your compliance team chooses which providers are allowed
Open AIPE
🚀
Pick your model
Routed through your own API keys — no lock-in
OpenAI (GPT)Cloud • US
Anthropic (Claude)Cloud • US
Google GeminiCloud • US/EU
xAI GrokCloud • US
Ollama — any local modelOn-prem • private
+ any provider with a chat-completions APIBYOM
📥 Upload XLIFF 🤖 150 seg batches 📚 Glossary 📤 XLIFF out
05Post-Processing • In dev

Automated cleanup before a human opens the file.

Linguists shouldn't waste time on problems a machine can catch first. The Post-Processing stage applies deterministic fixes that need no judgement: tag balance, number consistency with the source, leading / trailing space, consistent punctuation.

  • Tag validation — no dangling or mis-ordered placeholders
  • Number check — digits in source must match in target
  • Format cleanup — whitespace, smart-quote normalisation
  • Pre-linguist gate — flags what still needs human eyes
Preview Post-Processing
🔧
Automated checks
Run deterministically on every file
CheckSegmentsVerdict
Tag balance3,412✓ clean
Numbers match3,412⚠ 4 to review
Leading/trailing ws3,412✓ clean
Smart quotes3,412⚙ 27 normalised
06Linguist review • External

A professional linguist stays in charge — in any CAT tool.

AI drafts, humans decide. The cleaned bilingual XLIFF opens in whatever CAT tool the linguist prefers — memoQ, Trados Studio, XTM, Phrase, Wordfast, Smartcat — where they confirm, override, or rewrite the AI output with full TM leverage, concordance, and inline term highlighting. This step is outside locAGI by design: we don't lock you into any one vendor.

  • Any CAT tool that reads bilingual XLIFF works
  • AI output arrives as target, TM and TB still in effect
  • Concordance with previous project decisions
  • CAT-native QA runs alongside (Verifika, Xbench, built-ins)
Handled in the linguist's CAT
👨‍💻
Use your existing workflow
Bilingual XLIFF — any CAT tool
memoQ Trados Studio XTM Phrase Wordfast Smartcat OmegaT CafeTran
#SourceAI draftLinguist
142The patient must take the medication before meals.Der Patient muss das Medikament vor dem Essen einnehmen.✓ confirmed
143Consult your doctor.Konsultieren Sie Ihren Arzt.✎ edited
144Store at room temperature.Bei Raumtemperatur lagern.✓ confirmed
07QA Review — MQM

Score, flag and fix — with AI as your second pair of eyes.

After the linguist, a second AI pass reads the finished translation as a reviewer would. It flags potential issues using the industry-standard MQM framework — typed errors with severity 1-5, a suggested revised target, and a comment. You approve, override, or mark as false positive. Export as Excel and annotated bilingual XLIFF.

  • MQM error typology (Accuracy, Fluency, Terminology, Style, ...)
  • Severity 1–5 — weighted into an overall quality score
  • Multilingual view — 3+ languages side-by-side
  • Export Excel (per language) or bilingual XLIFF with inline CAT-tool comments
  • Model performance table — compare which AI caught what
Single-language QA Multilingual QA
🔎
Multilingual QA — quality 94.2 / 100
3 languages • one row per source ID
🌐 Merged 7 Copy.docx_ger 7 Copy.docx_por-BR 7 Copy.docx_fre-FR 7
ID
SOURCE
TARGET DE
TARGET PT-BR
TARGET FR-FR
2
Please alert the front desk.
Bitte informieren Sie die Rezeption.
Por favor avise a recepção.
Veuillez prévenir la réception.
3
https://svrehihelpdesk.hhsc.org/Ticket/16188
https://svrehihelpdesk.hhsc.org/Ticket/16188= SOURCE
https://svrehihelpdesk.hhsc.org/Ticket/16188= SOURCE
https://svrehihelpdesk.hhsc.org/Ticket/16188= SOURCE
5
Consult your doctor.
Fragen Sie Ihren Arzt.Acc·4
Consulte seu médico.
Consultez votre médecin.
🚫 False Positives 2 excluded from report
DE#17 · Client approved "Anzeichen"♻ restore
PT#58 · Informal register is intentional♻ restore
1
Info
2
Minor
3
Moderate
4
Major
5
Critical
08Client delivery

Hand off a bilingual file you trust.

The final XLIFF carries the linguist's confirmed translation, the QA reviewer's findings baked in as CAT-native comments (so they show up in memoQ, Trados, XTM — wherever the client opens the file), and the quality score as metadata. The Excel export gives the project manager the same findings in a format they can share, filter, and archive.

  • Bilingual XLIFF with inline CAT-native comments
  • Excel per language — linguist reply column included
  • ZIP package — all formats in one download
  • Full audit trail — who did what, when
See all activity
ACME_Q1_Pharma_DE.xliff
3,412 segments • Quality 94.2 / 100 • ready
XLIFF + comments Excel per lang ZIP package
3,412 segments reviewed
87 AI findings, 4 confirmed by linguist
Comments readable by any CAT tool
Export signed by reviewer & PM
09Platform

Built for production linguistics, not just demos.

Six things that turn locAGI from a clever tool into a platform you can run a real localisation business on.

⚙️
Technique

Chunked parallel inference

Every AI job is split into 150-segment batches. Each batch is a single LLM call with its own prompt, the linked termbase, and a running glossary harvested from earlier batches so terminology and tone stay consistent across a 10,000-segment file. Partial failures retry segment-by-segment — nothing is dropped silently.

150 seg / batch Streaming glossary Smart retry Deterministic output
🧠
Model-agnostic

Gateway to any AI model

Any provider, not a fixed list. locAGI is a gateway: plug in any provider that speaks a chat-completions API and it slots into every step. Swap models per project, per stage, or let the platform rank them by past performance on your domain. Your compliance team — not us — decides which providers are allowed for which data.

OpenAI Anthropic Google xAI Ollama (local / on-prem) + any chat-completions API
🧩
Extensible

Plugins & weekly improvements

Need a new rule, a new export format, a new MQM category? Plugin hooks for import, export, QA checks, and MT providers. Customer suggestions that benefit the whole platform are implemented free of charge, typically within a week. You're not paying twice for the roadmap.

Import / export hooks Custom QA checks Custom MT providers ~1-week turnaround
🔗
Collaboration

Give your linguist the fix, their way

Share findings with an external linguist via a secure signed link — they edit directly in the browser, no install, no account — or hand them the commented bilingual XLIFF + Excel package for offline work in their own CAT tool. Their edits flow back into the project automatically either way.

Signed share links Expiring tokens Offline XLIFF + Excel Round-trip merge
🔌
Automation

Full REST API

Every action the UI can perform is available over HTTP: create a project, upload an XLIFF, launch AIPE, poll QA status, fetch findings, trigger exports. Wire locAGI into your TMS, your CI, or your project-management tool.

REST & JSON API tokens Webhooks OpenAPI spec
☁️
Deployment

Cloud, private server, or local

Same product, three delivery modes. Use our managed cloud for fastest start; run it on your own private server when data has to stay inside your firewall; or install locally on a linguist's workstation for on-device privacy. Full feature parity in all three.

Managed cloud Self-hosted (Docker) Local install Air-gapped mode

Start at step 1, or jump straight to QA.

The pipeline is designed to be used end-to-end, but every stage stands on its own. Run an AI-only project with AIPE, use just the QA module on a file someone else translated, or drive the whole flow from termbase to delivery.