Pickup Director - AI-Powered Audiobook Proofreading on UGREEN NAS

Turn narration mistakes into actionable corrections, without leaving your NAS workflow.

Inspiration

Recording an audiobook or educational course involves much more than reading a script aloud. Even experienced narrators can accidentally skip words, repeat sentences, or change phrases while recording.

Finding those mistakes requires someone to listen carefully while following the original manuscript. For independent narrators, author-narrators, and course creators, this proofreading process can be repetitive, time-consuming, and difficult to manage.

We wanted to explore a simple question:

What if your NAS could help proofread your recordings against the manuscripts already stored alongside them?

Rather than building another standalone transcription tool, we envisioned an AI-powered assistant that integrates proofreading directly into a creator's existing file-management workflow.

That idea became Pickup Director.

What It Does

Pickup Director is an AI Skill designed for UGREEN NAS that compares an audiobook or course-chapter recording against its approved manuscript and generates a structured pickup list - a list of passages that may need to be recorded again.

The proposed workflow:

1. Locate the source files: Find the audio recording and corresponding manuscript on the NAS.
2. Inspect and transcribe: Check the audio metadata and transcribe the recording using the NAS's configured AI model.
3. Compare against the manuscript: Identify supported differences, including skipped words, changed phrases, and detectable repetitions.
4. Generate actionable pickups: Quote the relevant manuscript and transcript passages so narrators know what to review or re-record.
5. Separate uncertainty: Place ambiguous results in a dedicated Check by ear section instead of presenting them as confirmed mistakes.
6. Save and track progress: After user approval, save the pickup list and update the recording's NAS file-manager tags.
7. Verify corrections: Re-check newly recorded pickups and track whether the corrected audio has been incorporated into the assembled chapter.

Pickup Director distinguishes between pickups-needed, pickups-recorded, and proofed-clean, ensuring that re-recording a correction is not confused with verifying the finished chapter.

How We Built It

We developed Pickup Director using UGREEN's UGOS Skill Creation Tool (uliya-skill-kit 0.7.3) and the documented AI NAS capabilities.

The Skill is organized around a primary SKILL.md file and five supporting reference documents covering comparison rules, disclosures, output formatting, validation, and usage guidance.

AI NAS Integration

Our design orchestrates three registered expert agents:

• file-agent: Locates NAS files, checks recording metadata, verifies saved outputs, and manages file-status tags using supported ug-filesystem and ug-filemgr capabilities.
• document-agent: Retrieves manuscript content from supported documents through ug-document qa.
• knowledge-agent: Performs audio transcription and saves generated pickup lists using supported ug-knowledge capabilities.

The workflow uses supported delegate --target calls rather than custom executable scripts or unrestricted filesystem operations.

Reliability and Human Oversight

A major design priority was avoiding unsupported conclusions.

Pickup Director uses multi-word passage anchors, text normalization rules, and conservative comparison criteria. When passages cannot be uniquely aligned, the system marks them as uncertain rather than inventing an error.

The workflow also separates reading from writing, requests user approval for changes, handles delegation states, and verifies saved files through read-back checks.

Current Development Status

We have completed the initial Skill package, supporting reference documentation, and a structured validation plan.

The official static checker is reported to pass with zero errors and zero warnings.

Actual execution on a UGREEN NAS - including transcription behavior, document-reading fidelity, permissions, and tagging - remains pending finalist-stage hardware validation.

Challenges We Faced

1. Distinguishing Narration Errors From Transcription Errors

Speech recognition systems do not always preserve everything a narrator says. They may omit stutters, combine repeated phrases, or misrecognize proper names.

That creates a challenge: a transcript is not necessarily an exact representation of the audio.

Our approach is deliberately conservative. Pickup Director reports differences supported by the transcript, while uncertain passages are marked for human review.

We also designed test cases specifically to examine whether the transcription system preserves repeated phrases and false starts.

2. Aligning Similar Passages

Audiobooks can contain recurring phrases, repeated dialogue, and similar paragraphs.

A naive comparison could incorrectly match one passage with another and generate false pickup requests.

We developed comparison rules using unique multi-word anchors and explicit uncertainty handling to reduce these false matches.

3. Designing a Safe NAS Workflow

Pickup Director does more than generate text. It also saves results and changes file-status tags.

We needed to distinguish between a pickup that has been identified, a correction that has been recorded, and a final chapter that has actually been checked.

This led us to design separate workflow states, approval requirements, and save-verification steps.

4. Accounting for Privacy and Model Configuration

Unreleased manuscripts and audiobook recordings may contain sensitive or unpublished creative material.

Although the Skill itself does not call external services, the NAS owner's configured AI models may use cloud processing.

We therefore designed the workflow to disclose this possibility, obtain consent before model-dependent processing, and recommend local model configurations where appropriate.

What We Learned

Building Pickup Director taught us that a useful AI agent needs more than a capable language model.

Reliability depends on understanding what the AI can verify, recognizing uncertainty, and preserving human control over important decisions.

We also learned how to structure a multi-agent workflow around UGREEN's supported expert-agent and CLI capabilities rather than assuming unrestricted access to files or system commands.

Perhaps the most important lesson was that an AI-generated result should not automatically become a trusted production status.

A corrected pickup recording is not the same as a fully proofed audiobook chapter. Modeling that distinction made our workflow more practical and accountable.

What's Next

If selected for Round 2, our priority will be validating Pickup Director on the provided UGREEN NAS environment.

We plan to:

1. Test recordings with deliberately skipped, changed, and repeated passages.
2. Measure detection quality, false positives, and transcription-related limitations.
3. Validate manuscript retrieval and passage alignment.
4. Exercise the complete save, approval, and tagging workflow.
5. Demonstrate re-checking corrected takes and verifying the final assembled chapter.

Our long-term vision is to make Pickup Director a dependable NAS-native production assistant for independent creators - helping them spend less time manually organizing proofreading corrections and more time creating.

Pickup Director isn't intended to replace human listening judgment. It's designed to make that judgment faster, more organized, and better supported by evidence.

Built With

• ai-agents
• ai-workflow-automation
• audio-transcription
• cli
• document-processing
• file-management
• llm
• markdown
• multi-agent-orchestration
• natural-language-processing
• speech-to-text
• ugos
• ugreen-nas
• uliya-skill-kit
• yaml