Revix — Translation-quality workbench by Logrus Global

What Revix does

Three regimes. One audit trail.

The translation-quality stack is fragmented: single-segment opinions live in spreadsheets, mass corpus updates and global terminology changes live in CAT tools that can't reason about morphology or context, and QE-model evaluation lives in research notebooks that nobody can reproduce. Revix replaces all three with one workbench whose outputs are reproducible by construction.

01 · Single-segment scoring

Paste a segment. Get an MQM classification right now.

For when a translator, reviewer, or PM has one (source, target) pair and wants an opinion immediately. Paste it in, pick the languages, and Revix returns an MQM error classification — severity-ranked across Accuracy, Fluency, Terminology and Style — plus a TAUS EPIC AIQE score, on the same metric TAUS sells to LSPs per segment.

MQM TQE Analytic (Logrus Global) — Per MQM Council research and publications.
Optional glossary upload — deviations are flagged as Terminology > Glossary noncompliance.
No signup — the demo path. Lowest-friction way to see what Revix's classifications look like.

revix.logrusglobal.com / demo · en-GB → cs-CZ

Source · en-GB

Sakharov Prize Winner Nelson Mandela meets with EP President Simone Veil in Strasbourg.

Target · cs-CZ

Laureát Sacharovovy ceny Nelson Mandela při setkání s předsedkyní EP Simone Veilovou v Evropském parlamentu ve Štrasburku.

Major · Mistranslation Minor · Style / register Glossary OK

0.78 TAUS EPIC AIQE MQM TQE: Major × 1, Minor × 1 · Acc. → Mistranslation

DEMO · NO SIGNUP REQUIRED · MQM CLASSIFICATION + TAUS SCORE IN ~6 SECONDS

02 · Translation Corpus Sanitation

Enterprise-grade AI sanitation and corpus alignment — without hallucinations.

Stop feeding legacy errors to your LLMs and your translation pipelines. Revix isolates exactly which segments need correction with deterministic rule-based filters, applies targeted AI execution only on the filtered subset, and surfaces the precise delta for human validation. The model edits surgically inside the designated terminology — surrounding context and inline tags stay byte-identical to source. Linguists no longer edit line by line; they validate a curated sample of the delta.

Zero-hallucination guardrails — targeted AI execution is scoped to the rule-filtered subset. Inline tags (<bpt>, <g>, <x/>) are preserved as Unicode placeholders the model cannot break; segments outside the filter stay byte-identical to source.
Full translation-memory lifecycle — global glossary overhauls, brand renames, compliance updates, morphological corrections (flexes / case / agreement). Multi-condition filters (AND / OR / contains / does-not-contain) compose precisely which segments are in scope.
The human moat — linguists validate a curated sample of the delta, not edit from scratch. Review time drops ~90 %. Concurrent reviewers protected by 30 s heartbeat segment locks and a 90 s sweep, with live exclusion as you type.
System-agnostic infrastructure — XLIFF, SDLXLIFF, MQXLIFF, TMX, two-column XLSX in; same format out, inline tags re-inflated. Revix doesn't replace your CAT ecosystem — it cleans the data that feeds Trados, Phrase, BureauWorks, Catmint, or anything else.

The bottleneck has always been review, not edit. A senior reviewer normalising a brand term across a 50,000-segment memory used to mean reading every flagged segment and editing by hand — roughly 5 days of work. With Revix the edits are already made; the linguist validates a curated sample of the delta in a 4-hour exercise. Review time drops by ~90 %, and a multi-week archive-update collapses into a single working day.

revix.logrusglobal.com / projects / TRADE-2022-80052 · Filtered subset · Delta review

Revix Translation Corpus Sanitation — filtered subset of segments with AI-generated deltas ready for linguist validation, inline tags preserved.

CORPUS SANITATION · FILTERED SUBSET · DELTA READY FOR VALIDATION · INLINE TAGS PRESERVED

03 · AutoQE Studio

Decide whether a QE model actually works on your content.

Upload Quality Scorecards (AMTA-QE European Parliament XLSX, QTLaunchPad / DFKI XML & HTML, DGT — auto-detected). Run TAUS EPIC QE rate-limit-aware with per-segment caching. Then read off the confusion matrix, PR curve, F1 sweep, triage gain, score distributions, calibration, and an Economics widget pinned to real per-character TAUS billing.

Multi-format scorecard parsing — EP XLSX, DFKI / QTLaunchPad XML, DFKI HTML, DGT. Auto-detected from file content. Most tools support one.
Rate-limit-aware QE runs — 2 concurrent calls, 500 ms gap, retry-on-429 with backoff. Per-segment caching so re-runs don't double-charge.
Snapshots, not live data — every saved analysis is reproducible by construction. Re-run gives the same numbers.
AI Conclusion + Process Recommendation — Gemini Flash writes the diagnostic and operational guidance, citing the actual bootstrap CIs and p-values.

revix.logrusglobal.com / projects / 80052 / analysis · Confusion matrix · t = 0.80

Revix AutoQE Confusion Matrix — TP, FN, FP, TN counts at threshold 0.80, with precision, recall, specificity and NPV row/column metrics.

AUTOQE · CONFUSION MATRIX · 248 / 248 SEGMENTS SCORED · BASE RATE 16.5%

The differentiator

Statistical confidence is built in — not an afterthought.

The most common QE-evaluation mistake is looking at AUC = 0.85 on 300 segments and concluding "great QE model" — when the 90 % bootstrap CI is actually 0.55..0.95 and a re-sample would put it in a different tier. Revix flags this automatically. Every analysis carries bootstrap 90 % CIs on AUC, AUPRC and peak F1, plus a permutation-null distribution and an empirical p-value. When a verdict is supported only by a wide CI on a small sample, the verdict softens itself to "(uncertain — wide CI)" or "(indistinguishable from random)."

90%

Bootstrap CIs on AUC, AUPRC and peak F1 — every analysis, every snapshot.

2k

Permutation-null replicates per evaluation, with empirical p-value.

4×

Scorecard formats parsed and normalised — EP, DFKI XML, DFKI HTML, DGT.

€0.0002/char

Real per-character TAUS billing, reconciled against an actual invoice — not a guess.

· Verdict softening, in practice

When the data warrants confidence, Revix says so. When it doesn't, Revix says that too.

Most tooling shows AUC and lets the user draw their own conclusions. Revix's AI Conclusion is calibrated against the bootstrap and permutation layers — it cannot describe a QE model as "a useful ranker" if the CI overlaps no-skill. AI conclusions are pinned to snapshots, not live data, so the diagnostic remains valid even after the dataset is extended.

revix.logrusglobal.com / · Analysis · P & R vs threshold

Precision-recall curve with monotone envelope, AUPRC, and balanced + risk-driven threshold recommendations.

PRECISION–RECALL CURVE · MONOTONE ENVELOPE · BALANCED + RISK-DRIVEN THRESHOLDS

Why Revix, not generic LQA

Built for client and LSP operations, as well as academic benchmarks — Revix connects them.

Most QE-evaluation tooling stops at "here is the AUC". Revix is built for the moment after — when an operations lead has to decide whether to bet a workflow on the model, write the SLA, and explain the decision to a customer.

Statistical confidence, automatic

Bootstrap CIs on AUC, AUPRC and peak F1, plus a permutation-null AUC distribution and empirical p-value. Verdicts soften automatically when the data doesn't warrant them. Rare in the market.

AI conclusions pinned to snapshots

Every AI analysis is reproducible. Re-running against the same snapshot returns the same numbers (modulo Gemini's small non-determinism). No "the AI said something different yesterday".

Honest about cost

The Economics widget uses real per-character TAUS billing (€0.0002 / char, source + target combined, with a per-source-word approximation as a toggle). Reconciled against an actual TAUS invoice — not estimated.

Multi-format scorecard parsing

AMTA-QE European Parliament XLSX, QTLaunchPad / DFKI XML & HTML, DGT. Auto-detected from file content. Bulk folder upload with per-file error log. Most tools support one format. Revix supports four.

Roles & access

Granular roles. L360 SSO. No standalone passwords.

Revix-Admin

Full access including downloads, snapshots, bulk re-runs, dataset synthesis, destructive admin actions.

Revix-User

Full operational access. Save, generate, download, edit. Excludes admin-only destructive actions.

Revix-Viewer

Strictly read-only across the AutoQE module. View every dataset, snapshot and AI conclusion. Suitable for stakeholders, auditors and external review.

Revix-Customer

Own scorecards only. Self-service evaluation against the customer's own translation memory or pilot data.

Review, Sanitize legacy memory, Score, Evaluate, decide on QE strategy. One workbench.