Inspiration
Software engineers often inherit legacy projects with scattered documentation, conflicting requirements, and unclear development histories. Understanding an unfamiliar system is difficult, but understanding why decisions were made, what was actually deployed, and which information can still be trusted is often even harder.
Project Archaeologist AI was created to address this challenge by transforming fragmented technical records stored on UGREEN AI NAS into evidence-backed project knowledge and actionable recovery plans.
What it does
Project Archaeologist AI is an evidence-grounded legacy software project recovery agent built for UGREEN AI NAS.
It follows three core principles: RECONSTRUCT → VERIFY → RECOVER.
- Reconstruct: Build historical timelines and trace architectural decisions from requirements, meeting notes, technical reports, and release records.
- Verify: Distinguish PROPOSED, APPROVED, IMPLEMENTED, RELEASED, and VERIFIED lifecycle evidence, without assuming that a deployment claim proves operational success.
- Recover: Generate prioritized engineering recovery missions with evidence references, risk explanations, and measurable completion criteria.
Additional capabilities include a Project Confidence Map (CONFIRMED, CONFLICTED, UNKNOWN), cross-document contradiction detection, source reliability analysis, and identification of missing project knowledge.
How we built it
The project was developed using the official UGREEN Uliya Skill Development Kit and its supported Expert Agent architecture.
The Skill coordinates three registered agents:
- file-agent: Discovers project files within user-authorized NAS directories and records file metadata.
- document-agent: Extracts engineering evidence, reconciles documents, reconstructs architectural evolution, and prepares reports.
- knowledge-agent: Audits evidence consistency, contradictions, source independence, and knowledge gaps.
The parent Skill uses delegate orchestration, with explicit permissions, evidence handoffs, and failure handling. Source documents are treated as untrusted data.
To support evaluation, I created a synthetic legacy inventory management dataset containing seven interconnected historical documents, including conflicting requirements, ambiguous deployment records, and incomplete operational knowledge.
The Skill passed the official Uliya static validator with zero errors and zero warnings.
Challenges we ran into
The primary challenge was designing an AI workflow that distinguishes documented claims from verified engineering outcomes.
For example, a release note may indicate that a service was deployed, but this does not prove that its feature flag was enabled or that it successfully handled production traffic.
Other challenges included distinguishing planned and actual dates, detecting genuine contradictions without confusing normal version evolution, preventing fabricated citations, and handling missing or inaccessible documents.
These challenges were addressed through evidence-grounded reasoning rules, explicit lifecycle classifications, deterministic agent handoffs, and conservative uncertainty reporting.
Accomplishments that we're proud of
- Designed a structured three-agent project recovery workflow using official UGREEN capabilities.
- Developed architecture evolution analysis and five-stage project state recovery specifications.
- Created an Evidence Reliability Engine and Project Confidence Map.
- Designed a Recovery Mission Planner with prioritized, verifiable engineering tasks.
- Built a seven-document synthetic evaluation dataset, ground truth, and a 100-point evaluation rubric.
- Achieved
static-passwith zero errors and zero warnings.
What we learned
This project reinforced a central lesson in AI engineering: reliable AI must distinguish evidence, inference, and uncertainty rather than simply producing confident answers.
I also gained experience in multi-agent orchestration, tool boundary management, source attribution, permission-aware workflows, and designing measurable AI evaluation criteria.
What's next for Project Archaeologist AI
The current submission is a pre-test Skill specification that has passed official static validation. It has not yet undergone real NAS runtime testing.
If selected as a finalist, the next steps are to deploy it in the assigned UGREEN AI NAS environment, test real Expert Agent execution, measure results against the evaluation dataset, refine failure handling, and record a complete project recovery demonstration.
The long-term goal is to help engineers safely recover years of fragmented technical knowledge and confidently take over legacy software projects.
Built With
- multi-agent
- nas
- ugos
- ugreen
- uliya
- workflow
