Nextria offers CastleGuard AI, a secure on‑premise AI runtime platform that lets government, defence and regulated enterprises run AI applications on any hardware—from data‑center servers to edge devices—without vendor lock‑in or cloud dependence. The platform includes sovereign language models such as Evia for offline speech‑to‑translation and the Targeted Knowledge Model, a compact, auditable 1.3 B‑parameter model trained on open Canadian data, enabling sensitive workloads to maintain full data control and compliance.
Funding
Funding not disclosed
Founders
Product
Problem
Government, defence and regulated enterprises need AI capabilities that can operate on classified, air‑gapped or low‑connectivity networks, but most commercial generative models require cloud access, large compute resources and expose data to external providers, creating security and sovereignty risks.
Solution
Nextria provides CastleGuard AI, a secure on‑premise AI runtime platform that abstracts deployment so AI applications can run on any hardware—from data‑center servers to edge devices—without vendor lock‑in. The platform includes sovereign language models such as Evia, a Mixture‑of‑Experts transformer that delivers speech‑to‑translation, speaker identification and transcription entirely offline on mobile devices, and the Targeted Knowledge Model (TKM), a compact 1.3 B‑parameter small language model trained on open Canadian government data for transparent, auditable use. By keeping models and data within the customer’s infrastructure, Nextria enables sensitive workloads to benefit from modern machine‑learning functionality while maintaining full control over data, compliance and security. The runtime supports any AI application, including operational intelligence, real‑time translation, and document analysis, and can be deployed on‑premise, at the tactical edge, or on‑device without reliance on external cloud services.
Target Audience
Primary customers are Canadian federal, provincial and municipal government agencies, defence and security organizations, and regulated enterprises that require on‑premise or edge AI for classified, multilingual, or low‑connectivity operations.
Features
- CastleGuard AI runtime abstracts hardware and environment, allowing AI apps to run on‑premise, on edge servers, or directly on mobile devices
- Evia on‑device Mixture‑of‑Experts transformer reduces compute and memory by 75% while delivering real‑time speech recognition, speaker identification and bilingual translation without internet connectivity
- Targeted Knowledge Model (TKM) is a 1.3 B‑parameter MoE model (~3 GB FP32) that runs on modest GPU resources and can be fine‑tuned on internal datasets for secure, closed‑network use
- Fully transparent training on open Canadian government data ensures model sovereignty and auditability
- Supports French‑Canadian language model optimized for Canadian terminology and Indigenous language collaborations, keeping linguistic data local
- No vendor lock‑in; models and inference pipelines can be deployed in classified, air‑gapped, or resource‑constrained environments