đź’ˇ Inspiration
In today's digital landscape, professionals, researchers, and organizations handling sensitive records—whether legal contracts, medical reports, R&D intellectual property, or confidential financial sheets—face a critical dilemma. Cloud-based AI tools offer incredible productivity boosts, but they require sending private, legally protected data to external third-party servers, posing severe privacy risks and regulatory compliance violations (such as GDPR or professional secrecy).
We were inspired by a simple question: What if we could bring the power of autonomous multi-agent AI directly to the user's hardware, ensuring 100% data sovereignty without ever sacrificing analytical depth? The UGREEN AI NAS platform, powered by UGOS, presented the ideal decentralized ecosystem to turn this vision into reality.
⚙️ What it does
SovereignVault AI is a modular, 100% local multi-agent orchestration skill designed natively for UGREEN AI NAS. Instead of relying on a single generic prompt or cloud-dependent retrieval-augmented generation (RAG) pipelines, SovereignVault deploys a deterministic 4-stage multi-agent workflow directly on-premise:
- Ingestion & Security Guard Agent (
guard-agent): Validates folder access permissions, enforces zero outbound network connectivity (strict_offline_check), and extracts raw text locally from heterogeneous files (.pdf,.docx,.txt,.csv,.md). - Structured Information Extractor Agent (
extractor-agent): Processes text chunks through lightweight quantized local LLMs (such as Qwen-2.5 or Gemma via Ollama/llama.cpp) to extract key entities, dates, metrics, and core obligations dynamically tailored to the user's analysis focus. - Cross-Document Consistency & Anomaly Auditor Agent (
auditor-agent): Cross-references extracted facts and figures across multiple files in the target folder, flagging contradictions, missing sections, or numerical discrepancies. - Executive Synthesizer Agent (
synthesizer-agent): Compiles findings into an actionable markdown report saved directly inside the NAS shared folder, while appending an immutable entry to an encrypted local SQLite audit log (vault_history.db).
🛠️ How we built it
SovereignVault AI was engineered following the Uliya Skill Development Standard required for the UGREEN Hackathon:
- Core Orchestration: Built using a modular Python agent architecture structured cleanly around the
SKILL.mdspecification. - Local Inference & Parsing: Integrated local open-weights quantized models running via localhost API endpoints, paired with robust Python document-parsing libraries (
pdfplumber/PyMuPDF) for reliable local text extraction. - Secure Storage & Audit Trail: Leveraged native UGREEN NAS encrypted shared folders, local Docker/CLI runtime isolation, and embedded SQLite for tamper-evident local execution logging.
🧗‍♂️ Challenges we ran into
- Balancing Inference Footprint and Accuracy: Running multi-agent orchestration locally on NAS hardware required optimizing prompt structures and selecting highly efficient, quantized small language models (SLMs) to ensure fast execution without hitting memory bottlenecks.
- Ensuring Zero Outbound Traffic: Designing the strict offline validation layer (
strict_offline_check) to guarantee that agents operate in a fully air-gapped manner while still interfacing smoothly with the local UGOS runtime environment.
đź“š What we learned
Building SovereignVault AI reinforced our belief that the future of enterprise and professional AI is local, sovereign, and agentic. Hardware platforms like the UGREEN AI NAS bridge the gap between heavy cloud compute and complete data privacy, proving that advanced multi-agent intelligence does not require external servers to deliver professional-grade results.
