Inspiration
Creative studios, legal practices, and privacy-conscious enterprises produce vast archives of confidential contracts, financial ledgers, and proprietary media stems. Modern AI promise requires uploading these sensitive assets to public cloud vector APIs (OpenAI, Anthropic, Pinecone)—creating unacceptable surveillance risks, NDA violations, and copyright leaks. We built Private Vault AI to bring sovereign, air-gapped intelligence directly onto private UGREEN NAS hardware with zero cloud telemetry egress.
What it does
Private Vault AI is a turnkey, containerised UGREEN UGOS Skill that autonomously monitors NAS shared storage folders, extracts and chunks text from dropped documents (PDF, Markdown, TXT, DOCX), computes vector embeddings locally, and serves an intuitive natural-language query interface with exact document citations. All processing and storage remain 100% on-premises.
How we built it
- UGOS Integration: Leverages UGOS Storage Manager for direct volume mounts and filesystem event monitoring.
- Local Inference Engine: Quantized Llama-3.2-3B-Instruct running locally on-device.
- Embedded Vector Database: SQLite-vec / embedded Qdrant for fast semantic similarity retrieval without remote cloud databases.
- Zero-Egress Security: Strict container network policies preventing any external telemetry or data leakage.
- Standardised Uliya Skill: Fully compliant with the Uliya Skill Development Standard (SKILL.md specification with air-gap verification procedures).
Challenges we ran into
Balancing fast semantic search latency with the hardware footprint of home and SMB NAS devices required careful selection of quantized model weights (4-bit quantization) and efficient local embedding models that maintain high precision without exhausting NAS system memory.
Accomplishments that we are proud of
- Delivering sub-1.5 second retrieval latency for complex contract queries entirely on local hardware.
- Creating a zero-cloud, zero-subscription knowledge engine that operates with complete data sovereignty.
- Standardising the packaging into an easily distributable Uliya Skill package ready for the UGREEN ecosystem.
What we learned
Local NAS hardware equipped with modern multi-core processors and NPU/iGPU capabilities is more than capable of running production-grade semantic search and agentic workflows when properly containerised and optimised.
What is next for Private Vault AI
- Hardware validation and benchmarking on the UGREEN iDX6011 Pro NAS during Round 2.
- Multi-modal local indexing for architectural blueprints, raw audio stems, and design assets.
- Turnkey 1-click installation package via the UGREEN UGOS App Center.
