Inspiration
Small and medium businesses, freelancers, and household CFOs manage a continuous influx of financial documents—ranging from recurring SaaS invoices and hardware receipts to vendor agreements. However, they face three persistent problems:
- Cloud Privacy & Compliance Vulnerabilities: Uploading balance sheets, pricing structures, and signed NDAs to third-party public cloud LLMs introduces serious data leak risks and regulatory non-compliance.
- Silent Capital Drains: Duplicate billings, unannounced price hikes, and automatic subscription renewals easily slip past manual bookkeeping.
- Information Fragmentation: Invoices and contracts remain trapped inside unindexed PDFs across local NAS shared directories.
We built DeepVault to transform the UGREEN AI NAS into a self-hosted, privacy-first compliance auditor that reconciles records deterministically right where data lives—without a single byte of private financial data ever leaving the local premises.
What it does
DeepVault is an autonomous orchestration skill designed for the UGREEN Uliya engine. When triggered via natural language (e.g., "Audit vendor invoices for Q3 and check renewal deadlines"), it executes a multi-phase audit workflow:
- Local Document Ingestion: Scans designated NAS directories for financial files via
file-agent(ug-filemgr), enforcing strict read-only intake to protect original documents from alteration. - Zero-Cloud OCR & Parsing: Extracts vendor names, invoice identifiers, issue dates, due dates, line items, and balances using
document-agent(ug-document). - Deterministic Arithmetic & Duplicate Auditing: Executes deterministic mathematical cross-checks ($\text{Subtotal} + \text{Tax} = \text{Total}$) to eliminate LLM calculation hallucinations. Runs rule-based heuristics across rolling 14-day time windows to catch duplicate charges and flags contract renewal deadlines (Urgent $\le$ 30 days, Attention 31–60 days, Monitored 61–90 days).
- Graceful Degraded Fallback: If a scanned invoice is skewed, corrupted, or low-resolution, DeepVault flags the item as
FLAGGED_MANUAL_REVIEWwithout guessing numbers or halting the batch. - Strict Human-in-the-Loop Safeguards: Enforces a strict
waiting_approvalstate before committing file changes, presenting an interactive summary matrix (Vendor, Invoice Number, Due Date, Total, and Discrepancy Status) for explicit user review. - Audit Ledger Generation: Persists structured, human-readable compliance audit reports in Markdown with standardized YAML frontmatter directly to
/nas/Finance/Audited/YYYY/YYYY-MM-DD_[audit_slug].md.
How we built it
DeepVault was built strictly following the Uliya Skill Standard v2.0.0 and UGREEN SkillHub submission guidelines:
- Declarative Orchestration: Implemented in
SKILL.mdusing formal delegation syntax (delegate --target <agent> --instruction '<task>') without prohibited executable binaries or shell scripts. - Capability Evidence: Mapped public CLI tools directly to registered child experts (
ug-document→document-agentandug-filemgr→file-agent). - Modular Progressive References: Standardized deterministic heuristics and report schemas inside
references/extraction-rules.mdandreferences/audit-report-template.md. - Finite-State Machine (FSM) Compliance: Explicitly handles all 5 lifecycle states (
completed,failed,cancelled,waiting_user, andwaiting_approval). - Automated Audit Pass: Packaged and validated with
check-skill.mjsinside an isolated Node.js container, achieving a cleanstatic-pass(0 errors, 0 warnings).
Challenges we ran into
- Zero-Hallucination Enforcement in Calculations: Standard probabilistic LLMs frequently hallucinate tax arithmetic or round numbers arbitrarily. We resolved this by separating LLM entity extraction from mathematical verification: figures are reconciled through strict arithmetic equations ($\text{Subtotal} + \text{Tax} = \text{Total}$), and low-confidence scans are isolated into an unparseable review table rather than interpolated.
- Strict Mechanical Linter Rules: Conforming to Uliya’s mechanical validation checks required aligning level-1 Markdown headings (
# Goal,# Required input,# Generation-time capability evidence,# Workflow,# Completion and failure) and precise Unicode arrow formatting (- \` → ```). - Air-Gapped Data Safety: Ensuring that document parsing and audit compilation operate strictly within the on-premise UGOS environment (
SECURITY_DIRECT_FILE_ACCESScompliant) with zero external network data transmission.
Accomplishments that we're proud of
- Achieving a verified
static-passaudit result (0 errors, 0 warnings) on the official UGREEN Uliya test suite. - Proving that complex enterprise financial reconciliation can run completely on-premise without a single byte of confidential financial data leaving the local network.
- Building a deterministic audit engine that eliminates LLM calculation errors while maintaining user-friendly conversational invocation.
What we learned
- How Uliya effectively isolates declarative orchestration from domain-specific agent execution across UGOS.
- Structuring strict Human-in-the-Loop approval checkpoints to safely execute high-impact storage operations.
- Best practices for optimizing document OCR heuristics on local, power-efficient NAS hardware.
What's next for DeepVault: On-Premise Financial Auditor
- Round 2 Hardware Testing: Benchmarking multi-page OCR throughput and memory allocation directly on the physical UGREEN AI NAS iDX6011 Pro environment.
- Multi-Currency Reconciliation: Adding local exchange rate tables for automated cross-border accounting.
- Self-Hosted ERP Sync: Exporting RFC-compliant transaction CSVs for Firefly III and Akaunting.
- Ambient UGOS Push Alerts: Delivering proactive reminders to UGOS mobile apps 7 days prior to critical contract renewals.
