An overview of bAIbel AV

bAIbel AV is a desktop application for translating documents with the help of AI. At its core it is a familiar computer-assisted translation tool. Around that core it adds a small number of capabilities that exist because the engine now doing much of the work is a large language model. This page describes what the application does and what is new about it. The ten guides that follow cover how to use each part.

Figure 1. The application at a glance. A familiar translation workflow at the centre — import, translate, export — with a confidentiality layer, a choice of AI model, and human-reviewed AI assistance around it.

What will feel familiar

If you have used a computer-assisted translation tool, the shape of bAIbel AV will be recognisable. The foundations of the craft are all present, and they behave as you would expect.

Import and export. It reads the common professional formats — Word, PowerPoint, and the XLIFF family used to move work between tools — and writes the finished translation back to its original format.
The editor. A segment grid with source and target side by side, progress tracking, and a view that exposes inline codes and tags when formatting needs placing.
Translation memory. Your current text is matched against past translations, with fuzzy scoring, insertion, and write-back.
Termbases. Approved terminology is surfaced as you work, and you can add to it.
Alignment and extraction. Existing translations can be turned into memory, and existing documents into termbases.

In its bones, it is a standard tool for the job.

What is new, and why it matters now

Four capabilities set bAIbel AV apart. Each exists for the same reason: translating with AI means sending text to a model, and that single fact changes the requirements of the work.

Confidentiality before the model sees the text

Machine translation sends your text to a model run by someone else. For material that is already public, that is fine. For a contract, an unpublished earnings release, or client work under non-disclosure, it is not. bAIbel AV can hide the sensitive parts — names, numbers, standard clauses — before anything is sent, and restore them afterwards. The masking is reversible, and you can test that the originals come back before you rely on it. This is the application’s central idea: the help of a capable model without handing it your confidential content.

Your choice of model, and what it means for privacy

bAIbel AV is not tied to one AI provider. You connect the model you choose, and you can use a different one for each project, or even for each task. The application is plain about the trade-off: the most capable models are run by large companies, others can be hosted under stricter terms, and some you can run entirely on your own hardware. Because the model landscape changes quickly and confidentiality needs vary from one job to the next, it is this choice — rather than a single fixed engine — that lets you match the tool to the work.

Expert-in-the-loop: AI proposes, the translator decides

When the application uses AI to make changes across a translation, it does not apply them silently. It proposes each change; you review and validate; only what you accept reaches the file. This matters because an ordinary find-and-replace cannot understand grammar. Changing a term in German, French, Spanish, or Russian ripples through articles, agreement, and case endings in ways a blind replacement gets wrong. An AI agent handles that correctly. Keeping the translator in the loop — the industry specialist, not the model — keeps judgement and voice in charge of the result.

Document understanding for machine translation

Rather than translate one sentence at a time in isolation, bAIbel AV can read a document’s structure — its sections, their purpose, and a summary of the whole — and use that as context for more consistent results. It also lets you choose a strategy to fit the document: adapt closely-matching memory for a revised text, or translate afresh for a new one.

A related point is worth stating plainly. The instructions that drive the AI are open to you in a prompt editor, rather than hidden inside the program. Transparency of this kind is part of the same principle as human review: you can see, and shape, how the tool behaves.

The throughline

None of this removes the translator. It takes on the parts of the work that machines do well — the lookups, the bulk edits, the first drafts — protects the parts that must stay confidential, and leaves the judgement where it belongs. Two ideas run through the whole application: keep confidential content confidential, and keep the translator in control.

The functions, guide by guide

Guide	Function
1. Core flow	The translation workflow from import, through memory and terminology, to export — with confidentiality built in.
2. Sectionizing	Machine translation with document context, and choosing a strategy to match the document.
3. The two views	Plain text for translating, and a formatting view for placing codes and tags.
4. Agentic editing	Reviewed, grammar-aware changes across a whole translation.
5. Choosing a model	Connecting and selecting the AI provider that fits each job and its privacy needs.
6. Numerical obfuscation	Hiding numbers before the model and restoring them afterwards.
7. PDF preparation	Converting PDFs into clean, translatable text, with a choice of converter.
8. Terminology extraction	Building termbases from documents, in one language or two.
9. TM alignment	Turning past translations into reusable translation memory.
10. Configuration	Connecting outside services, shaping the AI’s instructions, and building reusable patterns.