AI-Native Translation Platform

AI agents that translate,
review, and QA for you.

Autonomous agents run entire jobs. LLMs work as translation engines. AI chat sees every segment. You stay in control — step in anytime.

TilBridge TMS product-guide.xliff EN → DE
Translator agent completed · 118 of 127 segments translated Done
# Source Target
1 Getting Started
Erste Schritte 101 TM
2 Click {1}Save{/1} to confirm.
Klicken Sie auf {1}Speichern{/1}. Agent
3 Your changes have been saved successfully.
Ihre Änderungen wurden erfolgreich gespeichert.| Agent
4 Maximum file size: 50 MB
Maximale Dateigröße: 50 MB Agent
5 Click here to continue.
Klicken Sie hier, um fortzufahren. Agent
Confirmed: 118 / 127
93%
1,284 words · 7,821 chars
3
AI roles in the workflow
8
TM leverage bands
21
Document formats
100%
Traceable edits

Capabilities

AI-native from day one

AI Roles

A terminologist, a translator and a reviewer — each one a workflow step. Assign a role and it runs the moment its step is ready.

Step 2 translator
TM band → MT draft → LLM post-edit
actor: agent auto on prev submitted

Project Chat

Ask about one job or a whole project. The assistant reads segments, TM matches, terminology and QA reports, and answers with sources.

LLM Translation

Use an LLM as a translation engine alongside classic MT. Tags preserved, glossary applied, cost recorded.

AI Match Repair

Soon

When a TM match is close but not perfect, AI adapts it to the changed source. Minimal edits, maximum precision.

Reference Materials

Soon

Upload style guides, briefs and past translations so AI roles and chat always carry the project context.

Translation Memory

Exact and context matches, fuzzy search with word-level diffs, and 8 leverage bands feeding analysis and pre-translation.

Rich Text Editor

ProseMirror-powered, with inline tag handling, segment split and join, and full RTL and CJK support.

Quality Assurance

Checks gate segment confirmation: empty target, number consistency, tag integrity, resegmentation integrity. Job-wide runs are published and stored.

21 File Formats

Word, Excel, PowerPoint, OpenDocument, HTML, XML, JSON and IDML — extracted and rebuilt through Okapi with structure preserved.

Other CAT tools bolt AI onto

decades-old architecture.

We built AI into

every layer from day one.

Workflow — handbook-v4.docx
Running
1 Term extraction term_expert Completed

142 candidates → 38 approved into the term base

auto on previous completed
2 Translation translator Active
TM band MT draft + glossary LLM post-edit
264 / 428 segments 18.2K tokens $0.14
auto on previous submitted
3 Review + QA reviewer Pending
manual dispatch
4 Sign-off human Pending
revision #2841 actor: agent · translator

AI Roles

AI that works the job, not a chatbot

A terminologist, a translator and a reviewer. Each is a workflow step with its own resources and rules — assign it, and it runs the moment the step becomes ready.

Three roles

Term Expert builds the glossary, Translator translates or post-edits MT, Reviewer edits the text and runs QA.

Starts on its own

Chain a step to the previous one and the AI role starts automatically on submit or completion. A whole job can run from import to delivery.

The same resources as a human

Translation memory, term base and machine translation — the same attachments, the same access rights, the same audit trail.

Every edit is attributable

Each segment revision records which role wrote it. Token spend and cost are recorded per run.

AI Chat — product-manual.xliff

Is #12 consistent with #5? They both mention the same feature.

AI

Good catch. #12 translates "Translation Memory" as "Übersetzungsspeicher", but #5 uses "Translation-Speicher". Your term base has "Übersetzungsspeicher" as the approved term — I'd update #5.

Ask about your translation…
342 in · 89 out Session: 4 messages

Project Chat

Ask about the job — or the whole project

A conversational assistant scoped to a job, a set of jobs, or an entire project. Built for project managers and admins.

Real scope, not a guess

The assistant sees the job or project it was opened in, and only that.

Answers with sources

It names the segments and reports it used, so you can check the claim.

Streams as it goes

Answers appear as they are produced — no waiting for the full reply.

Conversations persist

Each conversation is yours and picks up where you left off.

Machine Translation — Segment 8
Source (EN)

Press the key to export your {1}document{/1}.

mt:deepl MT

Drücken Sie die Taste, um Ihr {1}Dokument{/1} zu exportieren.

ai:gemini LLM

Betätigen Sie die Taste, um Ihr {1}Dokument{/1} zu exportieren.

2 engines · 1.2s
All tags preserved

AI translation

An LLM as a translation engine

Gemini as an LLM engine and DeepL for classic MT — one interface, tag safety, glossary support and cost accounting. More providers are landing.

Tag integrity

Inline tags stay where they belong. Formatting survives the translation.

Any language pair

LLM engines handle practically any language. No per-pair setup.

Your keys, your account

Bring your own provider keys per organization. Keys are encrypted and verified before they go live.

Cost accounting

Token use and cost are recorded per call, per feature and per job.

Match repair

Soon

AI repairs an imperfect match

AI adapts a close-but-not-exact TM match to what changed in the source. Minimal edits, maximum precision.

Automatic repair

Runs on the first fuzzy match of a segment. AI can be switched on and off at any time.

Change highlighting

You see exactly what AI changed — additions, deletions and edits are highlighted.

Apply in one click

Accept the repaired version in one click — or keep the original.

Structure preserved

Tone, terminology, style and tags survive. Only what changed in the source changes.

TM Match — Segment 23
AI
TM Match 87%
Source diff

Click the buttonPress the key to export your reportdocument.

TGT

Klicken Sie auf die Schaltfläche, um Ihren Bericht zu exportieren.

TGT-AI

Klicken Sie auf die SchaltflächeDrücken Sie die Taste, um Ihren BerichtDokument zu exportieren.

2 changes 0.3s

Intelligence

MT, AI roles and search

Choose your engines, bring your own keys, and see what every call cost.

DeepL

Live

High quality for European languages, with provider-side glossary support

Gemini

Live

LLM-backed translation with context and glossary carried in the request

Google Cloud

Soon

Broadest language coverage

Tilmoch

Soon

Locally built engine from Uzbekistan

MangiSoz

Soon

Locally built engine from Kazakhstan

Self-hosted

Soon

Your own models behind an OpenAI-compatible API

Key capabilities

Tag-safe translation
Glossary terms carried into the request
Per-organization keys (BYOK)
Cost tracking per provider and job
Batch pre-translation

Observability & evaluation

AI output is only the start. Keep the evidence.

Connect translation results to the model, checks, and human feedback behind them. Build a clearer picture of what happened—and what deserves a closer look.

Follow the output back to its origin

Keep results connected to the model and configuration that produced them. With Langfuse enabled, inspect the exact prompt and response.

See what the checks actually establish

Distinguish verified term renderings, conflicting proposals, and committed review outcomes. A missing exact term match is not automatically a translation error.

Connect human feedback to AI work

When people confirm translations or review term proposals, retain the outcome alongside the original AI candidate. Autonomous work does not require human feedback to be evaluated.

In-app evaluation workspace

Soon

Illustrative evidence

From candidate to outcome

Three example records · different evidence, clearly labeled

  1. Translator

    Deterministic check

    Canonical term rendering found

    Exact rendering verified for the supplied term. Contextual suitability is a separate question.

  2. Term Expert

    Human feedback

    Proposal edited, then approved

    The review outcome is linked to the original AI proposal.

  3. Reviewer

    Committed outcome

    Revision committed

    The recorded result reflects writeback and validation—not just the model’s verdict.

Interface concept · illustrative data

The evidence foundation is implemented. This in-app view is coming soon.

Your investigation workspace. Your Langfuse.

Optionally connect your organization’s Langfuse project for full AI transcripts and selected evaluation scores.

TilBridge retains evaluation facts independently of Langfuse. Checks and human effort are distinct signals—not a single translation-quality score. Semantic judging remains experimental and uncalibrated.

Reference materials

Soon

Give AI the whole picture

Upload style guides, briefs, past translations, terminology tables or URLs. Documents are chunked, indexed and made available to AI roles and to chat.

Any document type

PDFs, Word files, terminology spreadsheets, URLs, past translations, annotated documents.

Automatic chunking and indexing

Documents are split into sections and indexed automatically. No manual markup needed.

Alignment of past translations

Upload a source and its translation as a pair. TilBridge aligns them — even when paragraphs do not line up one to one.

AI search by meaning

Roles and chat search reference material by meaning, not keywords. Relevant fragments surface in the CAT panel.

References — Legal Translation Project
4 documents
PDF

Client_Style_Guide_DE.pdf

Style Guide
DE 48 chunks
Indexed
DOCX

Previous_Contract_EN-DE.docx

Prev. Translation
EN, DE 124 pairs aligned
Indexed

client.example.com/legal-glossary

URL
EN 12 chunks
Indexed
XLSX

GDPR_Terminology_Multilingual.xlsx

Terminology
EN, DE, FR, RU 340 terms
Indexed
Reference search — “formal register for contracts”
0.94 Style Guide › Chapter 3 › Register Rules

"All contractual obligations must use the formal register. Use Vertrag not Deal. Avoid colloquial expressions in legal contexts."

0.89 Prev. Translation › Clause 4.2

"The controller shall implement appropriate measures." → "Der Verantwortliche setzt geeignete Maßnahmen um."

0.82 GDPR Terminology › Row 47

EN: data controller → DE: Verantwortlicher — Legal term, GDPR Art. 4(7)

Context summary AI-generated

Covers formality requirements for German contract language. Specifies formal register, approved legal terminology, and tone guidelines from the client's style guide.

Hybrid retrieval dense + sparse
524 chunks indexed
12ms

Editor

Built on how translators actually work

A CAT editor built from scratch — not a rich-text box with a translation panel bolted on.

Terminology control

Approved terms are underlined, forbidden terms struck through

QA on confirm

Checks run automatically — errors block confirmation until fixed or deliberately overridden

Number and tag checks

Catches mismatched numbers, missing tags and broken formatting

Split and join segments

Fix bad segmentation without breaking the document structure or the export

TilBridge — user-manual.xliff
QA issues found — press Ctrl+Enter again to override
# Source Target
5 Welcome to TilBridge TMS. Willkommen bei TilBridge TMS.
6 Update the database settings. Datenbankeinstellungen aktualisieren.
7 Maximum file size: 50 MB
Maximale Dateigröße: 500 MB
8 {1}Delete{/1} the erased records
Die alten Einträge entfernen
QA Issues 1 error 1 warning
E

#7 · number_consistency

Number "50" in source but "500" in target

W

#8 · tag_integrity

Tag pair {1}…{/1} present in source, missing in target

Jobs project-manual.xliff — Editor
EN → DE | 247 segments | 3,842 words
B
I
U
1 User Guide
Benutzerhandbuch 101 TM
2 Getting Started
Erste Schritte 100 TM
3 Click {1}Save{/1} to confirm your changes.
Klicken Sie auf {1}Speichern{/1}, um Ihre Änderungen zu bestätigen. 95 TM
4 The maximum file size is 50 MB. Die maximale Dateigröße beträgt 50 MB.|
5 Export your project as PDF or DOCX.
6 Contact support for assistance.
7 Version 2.1 Release Notes

Translation memory

Two ways to find the right match — a third on the way

Exact and context matches, plus fuzzy search with word-level diffs. Semantic matching is in development.

Translation Memory Panel
Current Segment

Save your changes

Exact match 100%
Match

Save your changes

Translation

Speichern Sie Ihre Änderungen

Hash Trigram Embed Auto
100%

Exact match

Finds identical source text instantly. Context matches (101%) take the surrounding segments into account.

87%

Fuzzy match

Finds similar segments even when the wording differs. Diff highlighting shows exactly what changed.

0.94

Semantic match

Soon

Search by meaning finds translations that mean the same thing, even when the wording is completely different.

Auto

Repetition autofill

Confirm a segment once and identical ones fill in automatically. Pre-translate on import: 101% confirmed, the rest left as drafts.

QA Results — project-manual.xliff
Segment 47 1 error
Source

Retain the records for 1,000.50 days.

Target

Bewahren Sie die Daten 1.500,50 Tage lang auf.

number_consistency: source 1,000.50 is not present in the target
Checks Run
empty_target
number_consistency
tag_integrity
resegmentation_integrity
Ctrl+Enter to confirm Blocked — fix errors first

Quality assurance

Catch errors before delivery

QA runs when a segment is confirmed. Errors block confirmation — fix them, or override deliberately. Job-wide runs are published and stored.

On confirm

Errors block confirmation. Press Ctrl+Enter again to override deliberately.

Whole-job runs

Run QA across every segment. Results are published and attached to each segment.

Four checks today

Empty target, number consistency, tag integrity, resegmentation integrity. Terminology and spelling checks are in development.

Locale-aware numbers

"1,000.50" (EN) matches "1.000,50" (DE) correctly. No false positives.

Search and navigation

Find what you already translated

Concordance across every attached translation memory, with matches highlighted. Global search, find and replace, and segment filters are in development.

Full-text search across all attached TMs. Past translations surface with the matching part highlighted.

certificate of origin | 4 results
All Segments TM TB Semantic
Certificate of Origin Ursprungszeugnis
The certificate of origin must be signed Das Ursprungszeugnis muss unterschrieben werden
Attach the original certificate of origin Fügen Sie das Original-Ursprungszeugnis bei
Document proving product origin Dokument zum Nachweis der Produktherkunft

Production and workflow

More than an editor

Revision history

Every edit is tracked with its origin — human, TM, MT or an AI role. The provenance survives for the life of the job.

Analysis and pre-translation

Run a leverage analysis across 8 bands before work starts, then pre-translate from TM, MT or both with configurable thresholds.

Workflow steps and chaining

Order the steps, assign a person or an AI role to each, and let a step start automatically when the previous one is submitted or completed.

Dispatch, accept, push back

Send work to a vendor, see them accept and submit it, and push it back with a reason when it needs rework.

Notifications

Ready work, push-backs, submissions and failed AI runs reach the right person in the app and by email.

Term base

A glossary of approved and forbidden terms, suggested as you translate.

Pivot translation

Route a language through a pivot: translate once into the pivot, then feed every downstream branch from that result.

Job splitting

Split a job across several performers mid-step with frozen ranges — and reverse the split losslessly if plans change.

Comments

Threaded comments on any segment, with author and timestamp.

Production management

Projects, clients, vendors — one place

A project holds its jobs, its client, its deadline and its resources. Vendors get their own work plane; managers see every step, who holds it, and what it cost.

Projects / Acme Corp Q1 / user-manual-v3.xliff EN → DE
Open Editor
Download
Workflow
Translation
Translator Agent
Review
Michael R.
Done
Advance
Rollback
Step 2 of 2 · Started Jan 15
File Overview
Segments
247
Repetitions
18 (6 groups)
Words
12,450
Characters
74,820
Analysis
Run Analysis
Leverage
67%
Weighted Words
4,109
101%
22%
100%
15%
Fuzzy
30%
New
33%
Translation Memories
Attach TM
Main TM
R/W Penalty: 0
Client Ref TM
Read Penalty: 2
Reference Materials
Upload
Style Guide v2.pdf
524 chunks
Previous EN→DE.xlsx
Aligned
247 segments · 12,450 words Last analysis: 2 hours ago

Why TilBridge

What makes us different

Existing CAT tools
TilBridge TMS
Architecture
Java/.NET desktop on a 2000s codebase
Built from scratch on a modern stack
Performance
Seconds for a fuzzy search on a large TM
Instant results even on large memories
Large documents
Slows down past 5K segments
Smooth scrolling with 10,000+ segments
External agents
Not available
Connect your own AI assistant to the system
Traceability
Partial edit history
Every edit carries its origin; every AI call is accounted for
TM search
Exact and fuzzy
Exact, context and fuzzy with word-level diffs
Tag handling
Basic tag protection
Inline tags with noise suppression at extraction and restoration on export
Source differences
None or basic
Word-level diff highlighting on every match
QA
Run manually, review a list
Runs on confirm and blocks errors; job-wide runs are published
File formats
15–20 through plugins
21 native formats out of the box
Word count
Vendor-specific counting rules
The GMX-V 2.0 industry standard, script-aware
Concurrent access
Edit locks, no shared TMs
Shared translation memories across all jobs and users
AI in the workflow
A plugin, or a separate tool entirely
AI roles are workflow steps with rights, resources and provenance

Technology

A modern foundation

Built for speed and reliability — starts fast, stays smooth, and never gets in the way.

Rust backend

Handles large translation memories without a hint of slowdown.

axumtantivysqlxtokio

Svelte 5 frontend

A responsive, fast interface — no lag, no loading spinners between screens.

svelte 5prosemirrortailwind v4vite

Platform and data

One database, strict tenant isolation, and document conversion in an isolated service.

postgresqlokapis3mcp
Soon

Desktop app

A lightweight desktop build for offline work is planned for a later stage.

tauriwebviewoffline

Free Trial

Try TilBridge TMS for Free

We're onboarding new users. Leave your details and we'll set up your workspace.

By submitting, you consent to Tilbridge Tech LLP (BIN 260440008140) collecting and processing these details to respond and arrange access, including storage and email handling by hoster.kz in Kazakhstan. Inactive enquiries are retained for 12 months after the last meaningful contact. Marketing is optional. Details and withdrawal: Privacy notice