TENGWAR: Private AI Platform
A private AI assistant that answers questions about your company's documents with cited sources, generates reports and automates tasks. It is installed on NVIDIA DGX Spark servers at your premises and does not send your data to any cloud AI.
Problem & Context
Many organizations would benefit from an AI assistant that knows their internal documents, but cannot send confidential or regulated material to cloud AI providers: GDPR, trade secrets and intellectual property protection rule it out. TENGWAR is installed on NVIDIA DGX Spark servers at the client's premises: documents are processed there, indexed in a private vector database and queried from the web, Windows, Mac, iPhone or Android, with every answer backed by citations to specific pages of the original documents. It only connects to the outside for the integrations the client turns on, mobile access over an encrypted private network, and software updates.
Responsibilities
- Founded the product and led design, engineering and operations end to end
- Designed the 17-phase document understanding pipeline that turns raw files into searchable, multilingual, citable knowledge
- Built the tool-calling AI agent: document and web search, business integrations and report generation in Word, PDF and PowerPoint, with approval before actions
- Delivered the web, Windows and Mac apps on a shared Blazor interface, alongside the iPhone and Android apps
- Designed the enterprise security model: role-based access control, document access levels, encrypted connector credentials, auditable logging and default-deny tool authorization
- Orchestrated the on-premises deployment in three editions, from one to three DGX Spark units, with a single deploy command and full observability
Architecture & Stack
- Single-tenant on-premises deployment on NVIDIA DGX Spark: TENGWAR on one server; Pro and Max add one or two inference servers for the chat model. No cloud AI provider is ever contacted
- A single tool-calling agent on Microsoft Agent Framework answers in chat; long investigations split the work across specialist agents in the background
- 17-phase ingestion pipeline: layout detection, OCR, image and table extraction, translation, chunking and indexing, automated end to end
- Hybrid retrieval (dense vectors, BM25, ColBERT and a knowledge graph) with cross-encoder reranking, Self-RAG and CRAG correction to keep answers grounded
- Web, Windows and Mac apps on a shared Blazor interface, plus mobile apps; real-time streaming of answers and document processing progress
- Multi-layer security: access levels enforced at search time, default-deny tool authorization, approval before actions and audit logs with sensitive data redacted
Outcomes
- Private AI assistant for organizations that cannot use cloud AI: documents, models and queries stay at their premises
- The 17-phase ingestion automates layout detection, OCR, translation and indexing of documents in 24 formats and several languages
- Hybrid retrieval delivers grounded, cited answers. Every claim is traceable to a source page
- The agent decides when to search documents, check the web or the ERP, or draft a deliverable, and asks for approval before acting
- Three editions, from one to three DGX Spark units, with a one-off licence and no per-user fees, viable for SMEs and public administration without an IT department
- Web, Windows, Mac, iPhone and Android apps with the same experience, reducing adoption friction across mixed teams
Product screenshots
Demo
Learn More
Official product page: tengwar.net ↗