VaultMind: Local Cross-Artifact Research & Drift Auditor

An offline Uliya AI Skill for UGREEN NAS that links research PDFs, raw CSV logs, and codebases—automatically compiling literature matrices and detecting data-to-document contradictions locally.

Comment

Inspiration

Local AI NAS appliances excel at private, always-on storage, yet researchers and engineers face a pervasive problem: broken artifact lineage and silent contradiction drift. When conducting research or building software on a shared NAS, papers (.pdf), experiment logs (.csv/.json), codebase parameters (.py), and draft reports (.md/.docx) are scattered across directories. When an experiment dataset updates or code parameters shift, draft manuscripts frequently cite obsolete metrics. Commercial cloud LLMs are unviable due to strict privacy, IP retention risks, and prohibitive file context limits. We built VaultMind to turn UGREEN AI NAS into an autonomous, zero-leakage cross-artifact research and consistency auditor.

What it does

VaultMind is an offline Uliya AI Skill running natively on UGOS that bridges research documentation, tabular data, and codebases:

  1. Autonomous Cross-Artifact Audit: Scans project directories and flags numerical discrepancies between narrative draft claims (e.g., accuracy, learning rates) and actual ground truth in CSV logs or Python AST constants.
  2. Zero-Touch Literature Matrix Compiler: Parses research PDFs, extracts structured metadata (Methodology, Datasets, Metrics, Limitations), and compiles a unified comparative matrix with exact page-level citations.
  3. Grounded Hybrid Querying: Delivers deterministic semantic and keyword search across local project vaults with strict source attribution, ensuring zero hallucinations.
  4. Zero-Egress Privacy: Operates 100% on-device on the UGREEN NAS with no external API calls, safeguarding unreleased drafts and sensitive research artifacts.

How we built it

  1. Deterministic Parsing Layer: Uses PyMuPDF for PDF document layout inspection, Python's built-in ast module for deterministic code parameter extraction, and streaming chunk profilers for large CSV/JSON logs.
  2. Local Hybrid Index: Employs SQLite3 combined with Full-Text Search (FTS5) and lightweight on-device vector embeddings, enabling instantaneous retrieval without heavy container overhead.
  3. Uliya CLI Orchestration: Wrapped cleanly as a modular Uliya Skill via SKILL.md, adhering to UGOS CLI conventions with structured JSON and Markdown reporting.

Challenges we ran into

a. NAS Memory & Resource Constraints: Running AI reasoning on low-power NAS architectures without exhausting RAM. We solved this by enforcing deterministic AST/CSV parsing first and reserving local SLM/LLM inference strictly for synthesis tasks. b. Handling Corrupted/Non-Standard PDFs: Scanned image-only PDFs initially broke standard text pipelines. We implemented graceful degradation with zero-token fallback warnings (OCR_REQUIRED).

Accomplishments that we're proud of

  • Designed an incremental SHA-256 caching watcher that only parses modified artifacts, keeping background CPU load minimal.
  • Eliminated cloud dependency completely—all cross-referencing and matrix generation happen locally on-device.

What we learned

We mastered building strictly constrained, privacy-first AI agents tailored for always-on NAS environments, orchestrating local CLI agents through the Uliya skill ecosystem.

What's next for VaultMind

  • Integrating automated LaTeX bibtex synchronization.
  • Adding multi-user collaborative conflict resolution directly within UGOS shared team folders.

Built With

Share this project:

Updates

Submission history