Inspiration
Modern event and wedding photographers shoot between 2,000 and 5,000 uncompressed RAW captures (200GB to 1TB+) per assignment. However, delivering final albums requires a grueling initial step: manual photo culling. Photographers spend days zooming into burst shots at 100% crop just to discard out-of-focus frames, blinking eyes, and shutter misfires.
While cloud-based AI culling software exists, it introduces severe friction:
- Network Bandwidth & Ingestion Bottlenecks: Uploading hundreds of gigabytes of uncompressed RAW files (CR3, ARW, NEF) takes hours or days on standard internet connections.
- Recurring Cloud Costs & Data Privacy: Subscription fees escalate with storage volume, and uploading unreleased client portraits to public clouds creates client confidentiality concerns.
- Local Storage Disconnection: Raw media already resides safely on high-throughput local NAS storage, making remote uploads redundant.
We built LensCull AI to transform the UGREEN AI NAS into an autonomous, on-premise photo culling workstation that triages high-volume RAW shoots directly where the files live—at multi-gigabit local speeds with zero cloud exposure.
What it does
LensCull AI is an autonomous, declarative orchestration skill for UGREEN AI NAS (UGOS) designed to streamline post-shoot intake:
- Zero-Upload Ingestion: Discovers uncompressed RAW directories directly on NAS storage via
file-agent(ug-filemgr), enforcing strict read-only intake to guarantee master camera files are never altered or deleted. - On-Device Vision & Facial Triage: Batches embedded previews through
media-agent(ug-video) in bounded sliding windows (15–20 images) to evaluate pupil sharpness, detect closed eyes/blinks, assess composition, and cluster high-speed burst sequences. - Deterministic Classification: Scores and sorts assets into three clean tiers based on clear quality heuristics:
- Accepted (
5 Stars / Green): Pin-sharp focus, genuine expressions, open eyes, best frame of a burst sequence. - Candidate (
3 Stars / Yellow): Clean alternate compositions preserved for client proofing. - Rejected (
1 Star / Red): Blinks, defocus, severe exposure clipping, and redundant burst frames.
- Accepted (
- Human-in-the-Loop Safeguards: Enforces a strict
waiting_approvalstate before modifying any metadata, presenting an interactive culling breakdown table to the photographer for explicit confirmation. - Dual Non-Destructive Outputs: Automatically generates curated smart albums via
photo-agent(ug-photo) in the native UGOS gallery and compiles an audit ledger markdown report with standardized YAML frontmatter into{target_folder}/Culling_Reports/.
How we built it
- Uliya Skill Standard v2.0.0: Developed strictly under the official UGREEN Uliya framework using declarative workflow graphs in
SKILL.mdwithout disallowed binaries or external script dependencies. - Native Expert Delegation: Mapped CLI capabilities to registered UGOS sub-agents:
ug-photo→photo-agentfor gallery indexing and curated album assembly.ug-video→media-agentfor bounded visual analysis and facial blink evaluation.ug-filemgr→file-agentfor directory scanning and audit ledger persistence.
- Modular Quality Specifications: Isolated scoring matrices and output templates inside
references/culling-rubric.mdandreferences/culling-report-template.md. - Finite-State Machine (FSM) Lifecycle: Comprehensive coverage of all lifecycle states (
completed,failed,cancelled,waiting_user, andwaiting_approval). - Verified Codebase: Achieved an audited
static-pass(0 errors, 0 warnings) against the officialcheck-skill.mjstest suite in an isolated Docker sandbox.
Challenges we ran into
- Hardware Memory Throttling: Attempting to triage thousands of RAW files concurrently risks exhausting system RAM. We resolved this by engineering a bounded sliding-window mechanism (15–20 items per batch) with partial state preservation.
- Sandboxed Path Escalation Rules: The Uliya runtime prohibits direct Linux root filesystem access (
SECURITY_DIRECT_FILE_ACCESS). We refactored all file persistence logic to execute strictly via relative paths grounded in the user-providedtarget_folder. - Artistic Intent vs. Defect Detection: Distinguishing between an accidental motion blur and an intentional slow-shutter panning shot or moody silhouette required engineering a
FLAGGED_CREATIVE_REVIEWexception pipeline so creative shots are never falsely rejected.
Accomplishments that we're proud of
- Eliminating cloud upload wait times entirely by processing photography culling on local NAS storage.
- Delivering a verified, linter-compliant skill package meeting UGREEN's highest security and formatting standards.
- Ensuring complete data sovereignty and peace of mind for wedding and commercial photographers handling sensitive client media.
What we learned
- The architectural efficiency of edge AI NAS: how combining local NPU/GPU compute with multi-terabyte internal bus speeds unlocks use cases that are impractical in cloud-only environments.
- Best practices in writing declarative multi-agent workflows with deterministic human-in-the-loop gates.
What's next for LensCull AI
- Round 2 Hardware Benchmarking: Profiling embedded JPEG decoding and batch facial tagging directly on physical UGREEN AI NAS iDX6011 Pro hardware.
- Sidecar
.xmpSync: Exporting industry-standard Lightroom and Capture One rating tags directly into sidecar metadata files for instant desktop editing import. - Subject-Specific Grouping: Integrating local face clustering to sort wedding photos automatically by key VIPs (Bride, Groom, Family, Guests).
