Built for the disconnected edge.
Azcuna Dynamics, also known as Azcuna, is a U.S. defense technology company focused on one hard problem: making foreign-language material usable in places where the cloud is not available and the data cannot leave. We build the software that reads, translates, searches, and reasons over that material on the hardware already in the room.
Nearly every modern AI product assumes a live connection to a data center. That assumption fails inside a SCIF, aboard an aircraft, at a forward location without connectivity, and anywhere a document is too sensitive to upload. We design in the opposite direction. Every model and every pipeline runs locally on commodity hardware, nothing is transmitted off-box, and updates arrive as signed packages carried in by hand.
BABEL
The core platform. Offline machine translation across 200+ languages with priority tuning for Arabic, Farsi, Russian, Urdu, and Chinese; entity and relationship extraction; hybrid semantic and lexical search; and a document-to-decision workflow that keeps every claim tied to its source passage. Runs on a 16 GB laptop.
Platform page →ELLMENT
The E-2D Large Language Model Entity. BABEL's retrieval and grounded-generation core retargeted to NATOPS procedures, tactics, and mission communications for E-2D Advanced Hawkeye crews. Cites the governing passage for every answer, refuses when the corpus does not support one, and runs with zero network egress. Proposed to NAVAIR under Navy SBIR Phase I.
Program page →We build first and propose second. BABEL was benchmarked before anyone asked for it. The ELLMENT prototype was running before the SBIR proposal was submitted, at private expense, so that the evaluation could start from working software rather than a slide. The recorded session on the ELLMENT page is that prototype, unedited.
We treat model provenance as a supply-chain question. Every model we ship is an open-weight release from a disclosed publisher, verified against the publisher's hash, pinned by version, and recorded in the audit log with every query it answers. We default to U.S.-origin models and say so when we evaluate anything else.
We write down what the software cannot do. Offline translation on a laptop is not the best cloud service on every language, and our capability pages say so. A system that abstains when it should is worth more to an operator than one that always answers, and we engineer for the abstention.
Offline by default
Every model and pipeline runs on the operator's hardware. Whether a network exists is not something the software can detect, and it never needs to.
Sovereign data
Documents, embeddings, indexes, and logs stay inside the perimeter. There is no copy elsewhere, no telemetry, and no license server.
Mission-driven
We build for the analyst with a seized phone and the flight officer with a NATOPS question, not for a demo. If it does not survive the field, it does not ship.
- Jun 2026ELLMENT Phase I proposal submitted to NAVAIR PMA-231 under Navy SBIR topic DON26BZ01-NV010, with a working prototype built at private expense ahead of award. Award decision pending.
- Sep 2026Recorded ELLMENT prototype session published: real output, page-level citation, explicit refusal of a fabricated brevity code.
- NextELLMENT Phase I award and operator-in-the-loop evaluation with the E-2D community. That is the objective the company is built around right now.
Azcuna is the company's short name; Azcuna Dynamics LLC is the legal entity. BABEL is the platform. ELLMENT, with two L's, is the E-2D program and is pronounced “element.” If you see any of these names on a proposal, a SAM.gov record, or a conference badge, it is us.
Founded Azcuna Dynamics to bring offline, sovereign language AI to defense and intelligence teams operating beyond the reach of the cloud. Named inventor on two granted U.S. patents and a former data-science researcher at the Harvard Kennedy School.