Platform · BABEL
BABEL: foreign-language intelligence that never leaves the device.
BABEL reads, translates, and searches foreign-language documents entirely on-device. It resolves the people, organizations, places, dates, and values inside them, maps how they connect, and keeps every claim tied to the passage it came from. No cloud, no connection, and no data leaving the perimeter.
Recorded session · real output
The real app, on a UN Security Council resolution.
Eighty-two seconds, unedited: the BABEL desktop app on a 16 GB MacBook Air, with all network access blocked by the operating system. The recording signs in, pastes the title and three paragraphs of the official Russian text of UN Security Council resolution 2815 (2026), and translates them on the laptop. BABEL reports how sure the model was, checks that every number and glossary term came through, triages the text and screens its names against sanctions lists.
Recorded 21 October 2026. The Russian text is the UN's own, from docs.un.org. The clicks and typing are scripted; every frame comes from the app's own window, at real speed. The model is U.S.-origin, 8 billion parameters, 4-bit, on the laptop's GPU. The NOTIONAL banner is part of the app's demonstration build.
Check it against the UN's own English.
Each paragraph as BABEL translated it in the recording, next to the official English text the UN published for the same resolution. Paragraph 12 names the peacekeeping force by its Russian acronym, ВСООНК; BABEL's glossary maps it to UNFICYP, and the report's check line shows the match.
Paragraph 2 · Russian, UN text
настоятельно призывает стороны и всех участников прилагать все усилия для поддержания активного взаимодействия с Генеральным секретарем и его аппаратом;
BABEL, in the recording
Urges all parties and all participants to make every effort to maintain active interaction with the Secretary-General and his staff;
English, UN text
Strongly calls upon the sides and all parties to deploy all efforts to engage actively with the Secretary-General and his team;
Paragraph 12 · Russian, UN text
выражает свою полную поддержку ВСООНК, постановляет продлить их мандат на дополнительный период, заканчивающийся 31 января 2027 года, заявляет о своем намерении продолжать внимательно следить за положением на Кипре и заявляет также о своей готовности проводить по мере необходимости обзор выполнения настоящей резолюции с учетом рекомендаций Генерального секретаря, содержащихся в докладах, которые будут представлены во исполнение пункта 24 настоящей резолюции;
BABEL, in the recording
Expresses its full support for UNFICYP, decides to extend its mandate for an additional period ending on 31 January 2027, declares its intention to continue closely monitoring the situation in Cyprus and declares its readiness to conduct, as necessary, a review of the implementation of this resolution, taking into account the recommendations of the Secretary-General contained in the reports to be submitted in accordance with paragraph 24 of this resolution;
English, UN text
Expresses its full support for UNFICYP, decides to extend its mandate for a further period ending on 31 January 2027, affirms its intention to continue to monitor the situation in Cyprus closely, and further affirms its readiness to review the implementation of this resolution as necessary, taking into account the advice of the Secretary-General in the reports requested in paragraph 24 of this resolution;
Paragraph 19 · Russian, UN text
вновь настоятельно призывает лидеров обеих общин согласовать план устранения минной опасности на Кипре и продолжать работу в соответствии с этим планом, устранив существующие препятствия для этой работы, о которых говорится в пункте 17 доклада Генерального секретаря (S/2026/8), в целях достижения скорейшего прогресса в деле разминирования 29 оставшихся предполагаемых опасных районов на этом острове;
BABEL, in the recording
Reiterates its urgent call to the leaders of both communities to coordinate a plan for the removal of mine danger on Cyprus and to continue work in accordance with this plan, removing existing obstacles to this work, as mentioned in paragraph 17 of the Secretary-General's report (S/2026/8), with a view to achieving the earliest possible progress in the demining of the 29 remaining assumed dangerous areas on this island.
English, UN text
Continues to urge the leaders of both communities to agree and continue a plan of work to achieve a mine-free Cyprus, and to overcome the existing barriers to this work as outlined in paragraph 17 of the Secretary-General’s report (S/2026/8), in order to make expeditious progress towards clearing the 29 remaining suspected hazardous areas on the island;
Run it here.
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
From document to decision
One document in. A network out.
BABEL doesn't just translate. It resolves people, organizations, locations, dates, and values from foreign-language intercepts and maps how they connect. Hover any node to trace its relationships.
The pipeline
Five stages, all on the device. Select a stage to see what it does.
Ingest/Documents, images, and media loaded from local storage or seized media. No upload, no indexing service.
Language coverage
Constraint 1
Nothing leaves the perimeter
Seized media and collected documents often cannot touch a network, and a cloud translation service is an exfiltration path by definition. BABEL runs every model, index, and store locally, with no telemetry and no phone-home, so it can operate on the far side of an air gap.
Constraint 2
Every claim has a source
An analyst has to be able to defend a finding. Every extracted fact and summary stays tied to the passage it came from, so any claim can be checked against the original document in one step.
Platform facts
- Deployment
- Fully air-gapped, on-prem
- Minimum hardware
- Modern laptop (16 GB)
- Acceleration
- Apple Silicon / NVIDIA GPU
- Languages
- 200+, priority-tuned
- Priority coverage
- Arabic, Farsi, Russian, Chinese, Urdu, Pashto, Korean
- Outbound network
- None
- Update channel
- Signed packages, offline
- Model origin
- U.S.-origin foundation models
Built on BABEL · ELLMENT
The same core, retargeted to E-2D mission crews.
ELLMENT takes BABEL's retrieval and grounded-generation core and points it at NATOPS procedures, tactics, and mission communications. Same offline architecture, same source attribution, a different corpus and a different operator.
Platform FAQ
BABEL is the Azcuna Dynamics offline multilingual intelligence platform: machine translation, entity and relationship extraction, and semantic search over foreign-language documents, running entirely on hardware you control. It is also the core that ELLMENT is built on.
Priority coverage spans Arabic, Persian/Farsi, Russian, Chinese, Urdu, Pashto, Korean, and more, with broad additional coverage across 200+ languages. The supported set is driven by mission need.
No. Models, indexes, and storage all run locally. There are no cloud calls, no telemetry, and no model phone-home. BABEL is designed to operate fully air-gapped.
A modern 16 GB laptop is the baseline, with Apple Silicon or NVIDIA acceleration. Throughput scales with the machine, but the footprint is kept small on purpose so BABEL deploys at the edge instead of requiring a data center.
Updated models and indexes are delivered as signed packages that install from removable media, with no connection required. You control when and what updates.
Every extracted fact and summary stays tied to its source passage, so an analyst can check any claim against the original document in one step. BABEL is built to assist analysts, not replace their judgment.
ELLMENT is BABEL's retrieval and grounded-generation core retargeted to NATOPS procedures, tactics, and mission communications for E-2D Advanced Hawkeye crews. Same offline architecture and the same source attribution, applied to a different corpus and a different operator.
Evaluating offline language AI?
We work with defense and intelligence teams operating in disconnected environments.