bAIbel TMS is in development. It is not on sale, and nothing here is an offer.
bAIbel TMS runs multilingual production: the work that arrives, the people who do it, the memories and terminology it draws on, and the machine translation that now sits in the middle of all of it.
What separates it from the platforms you have already been shown is where the decisions live. The route work takes is something you author, not something you pick from a menu. The rules about where your clients' material may go are enforced at the border rather than written in a policy document. And what a budget will allow is settled before a run starts, not discovered on an invoice.
Each client organisation gets its own installation, its own database, its own object storage. Not a tenant in a shared platform with your data separated by a column. Run it on our infrastructure or on yours — the same four images, upgraded at one version, with the operational commands built into the image that runs the service.
A workflow is a sequence you build from a published palette of stages — translate a section at a time, or a segment at a time; analyse the memory first; summarise the document; check the inline codes; verify the translation against its source; compose a QA report. Nineteen stages today, and a stage becomes available by being registered, so the palette and the running system cannot disagree.
Place a gate and the work stops there and waits for a named person.
No summarize stage means no summary — not a default one generated quietly. No terminology pre-analysis means the run translates without leverage, which is a legitimate thing to ask for when a client has no memory worth leveraging. Absence means absence, and the request to the AI model reads as though the field were not there.
Four facts about an AI model provider are checked before your client's material can reach it: its trust tier, the region it is registered in, whether its agreement is still in force, and whether the instrument reaches raw content. Your client's own rules and the deployment's floor are folded together and the strictest wins. A lapsed agreement stops permitting anything.
Everything else a data processing agreement says is invisible to the code, and we would rather tell you which four facts are mechanised than imply the software understands your contracts.
Set a budget to enforce and a run is priced, padded by a margin measured from that workflow's own history, and checked against what is left — before anything is spent. If the cost cannot be computed at all, because an AI model has no rate in force, the run is refused rather than admitted.
A budget cannot bound spend it cannot price.
A section is translated with the document around it: what the document is about, the style it is written in, what this section is doing in it, the terminology that applies to these very segments, the wording already settled in earlier sections, and the renderings your memories already hold.
Pronouns, terms and cross-references resolve against the document, because the document is what the AI model was given.
The summary, the brief, the project description, the cultural guidance — all of it is prose about confidential material, so all of it gets stand-ins before it travels, carried behind markers rather than written into the instructions in the clear. If one of those markers cannot be placed, the call is refused rather than sent.
Every system we know of treats the operator's own notes as trusted. This one does not.
The email names the run and says where to look. It carries no reason, because a failure reason can contain an exception message and a gate reason names what a detector found inside your client's document — and a mail server is not covered by any confidentiality decision you have made.
The message points at the run. The run tells whoever signs in everything.
A manager uploads it. Your client drops it off through their own portal. Or another company's software delivers it through an agent-facing door, where the bytes never travel through an AI model. All three converge on the same pipeline, and the record of what was delivered says which kind of party received it back — filed where only you can reach it, collected by someone who signed in, or taken away by a third party's software.
react and react-dom. Nothing else. It is the one bundle served to people outside your company, and its dependency list is a security decision rather than a convenience.
Some material has to reach a reader the source was never written for. Recompose extracts the document's meanings as a structured tree, audits that tree against the source for fidelity and completeness, weights every detail for a named audience, adapts the culture-specific items — and then generates fresh prose from the tree alone. The source's wording never reaches the generator.
It arrives switched off. An administrator opens it deliberately, on an installation that has decided it wants it.
Projects, tasks and assignment. Memories with context-aware matching over a real search index. Terminology in and out as TBX, MultiTerm or memoQ CSV. TMX that leaves in the dialect it arrived in. Seven built-in quality checks and your own rules alongside them. Budgets, reporting, users, roles and groups.
Unglamorous, and the part a system is eliminated on.
TMS is in late alpha and is being readied with a small number of organisations. If you run multilingual production and would rather shape a system than be handed one, tell us what you run today.
