Foreign text in. Intelligence out.
Pick a sample document below and watch the BABEL pipeline detect the language, translate it, extract the entities, and summarize. This walkthrough is a guided replay of the pipeline on three sample documents, Arabic, Russian, and Farsi, so it runs instantly in any browser. The stages, the ordering, and the outputs match what the real engine produces on an analyst's laptop with no network.
For real, unedited model output, scroll to the recorded ELLMENT session. For the engine running on your own document, request a live walkthrough.
تم تأكيد وصول الشحنة إلى ميناء جدة يوم الثلاثاء. سيتولى السيد خالد المنصور التنسيق مع شركة الأفق للنقل. التكلفة الإجمالية 45000 دولار.
The shipment's arrival at the Port of Jeddah was confirmed for Tuesday. Mr. Khalid al-Mansour will coordinate with Al-Ufuq Transport Company. The total cost is $45,000.
Demonstration using sample documents · runs entirely in your browser, no data transmitted
1 · Detect
The script and language are identified from the text itself. No lookup service is called; the classifier is a local model a few megabytes in size.
2 · Translate
The document is chunked by script, translated by a quantized open-weight model on the local GPU, and cached so a second view is instant. Names and identifiers are held consistent across chunks.
3 · Extract
People, organizations, places, dates, and values are pulled from the English text with the original passage kept alongside each one, so every entity can be checked against its source.
4 · Summarize and link
A grounded summary is generated from the extracted evidence, and the entities are linked into a graph the analyst can pivot through. Nothing in any stage leaves the machine.
The engineering behind each stage is written up in the offline machine translation and air-gapped RAG deep dives.
Every entity resolved, and how they connect.
Extraction does not stop at a list. BABEL links people, organizations, locations, dates, and values into a relationship graph an analyst can pivot through, with the source passage one click from every node. Hover any node to isolate its connections.
The ELLMENT prototype, unedited.
Fifty-five seconds captured from the live prototype on a laptop. It answers a carrier NATOPS question with a page-level citation, then refuses a fabricated brevity code instead of guessing. That is the same retrieval and grounded-generation core that powers BABEL search, and the refusal is the feature.
Watch the session on the ELLMENT page →What we show a program office.
A browser page cannot prove an air gap. A laptop with the radio off can. A live walkthrough takes about forty minutes and covers:
- Network cable out, Wi-Fi off, on a 16 GB laptop you can inspect
- Your document, if you bring one, or a 40-page mixed-script sample set
- Translation with a glossary you dictate on the spot
- A question the corpus cannot answer, so you can watch the system refuse
- The audit log, the model provenance record, and the signed update package