Crystal Hollis

Python application, applied AI & accessibility

AI-Assisted Metadata Wizard

An operational, human-in-the-loop desktop workflow that converts reviewed intake records into structured media metadata, supports accessibility alt-text drafts, and keeps optional AI keywords under human quality control.

  • Python
  • Tkinter
  • OpenAI Vision
  • ExifTool
  • IPTC/XMP
  • Accessibility
Try the synthetic demoDownload desktop package

01

Overview

The Metadata Wizard grew from an earlier metadata and keyword prototype into an operational desktop workflow used in Dedman College media post-production. It connects reviewed SharePoint intake, structured metadata automation, assisted accessibility outputs, Adobe Bridge quality control, and the media-archive workflow.

2,628 filesDocumented summer backlog through August 2, 2026
Two review stagesReviewed intake and final Adobe Bridge quality control
Non-destructiveExifTool writes metadata without recompressing pixels

02

Problem or research question

Professional image batches require consistent metadata across multiple standards, but manual entry is repetitive and vulnerable to omissions. Accessibility descriptions and image-specific keywords add value, yet AI output cannot safely become authoritative metadata without review.

Design question: How can automation reduce repetitive metadata work while preserving human ownership of descriptive, accessibility, and taxonomy decisions?

03

Context and role

I designed and implemented the Python application around the live Dedman College post-production workflow. My role included transferring reviewed intake records into the application, mapping the intake-to-archive process, defining field transformations, building the Tkinter interface, integrating ExifTool and optional vision-model calls, and performing final metadata review and editing in Adobe Bridge.

This public case study and source copy intentionally replace organization-specific templates, names, paths, records, and keyword lists with synthetic examples.

04

Approach and workflow

Production workflow: A team member prepares the initial metadata record in SharePoint. I transfer the reviewed record into Metadata Wizard, which parses the intake and applies structured metadata while generating optional assisted outputs. I then review and refine the resulting metadata in Adobe Bridge before the files move forward in the media-archive workflow.

SharePoint intakeTeam member drafts the record
Reviewed transferHuman-confirmed fields enter the tool
Metadata WizardStructured and assisted outputs
Adobe Bridge QCFinal human review and editing
Archive workflowApproved assets move forward
  1. A team member prepares the initial metadata record in SharePoint.
  2. I transfer the reviewed record into Metadata Wizard and confirm the parsed fields.
  3. The Wizard applies structured IPTC/XMP metadata and generates optional accessibility or keyword outputs.
  4. Controlled-vocabulary matches, exceptions, successes, and failures remain visible for review.
  5. I review and refine the resulting metadata in Adobe Bridge.
  6. Approved assets continue into the media-archive workflow.

05

Tools and technologies

  • Python
  • Tkinter
  • Threading and queues
  • OpenAI Responses API
  • Vision input
  • ExifTool
  • IPTC
  • XMP
  • Pillow
  • macOS Keychain
  • Markdown

06

Data and inputs

  • JPG, JPEG, PNG, and WEBP image batches.
  • Labeled intake fields such as headline, description, location, title, batch keywords, and job identifier.
  • Configurable XMP templates stored outside the application logic.
  • An optional approved keyword list used to classify AI suggestions.
  • An optional API key retrieved from macOS Keychain or an environment variable.

The interactive demo uses a completely synthetic community technology workshop record and performs no uploads, API calls, or file writes.

07

Results and measurable outcomes

The Metadata Wizard is currently used in Dedman College media post-production as an intermediate automation and quality-control step between SharePoint intake and final Adobe Bridge review. It reduces repetitive field entry while preserving human review at both the intake and final metadata stages.

Operational scale and verified behaviors

  • The workflow supported a documented summer production backlog of 2,628 files across items scheduled through August 2, 2026.
  • Metadata-only writes preserve image pixels instead of recompressing the source asset.
  • Accessibility drafts are written to the dedicated XMP accessibility field and can be skipped when an existing value is present.
  • Optional AI keywords remain in a Markdown review report rather than being automatically committed to image files.
  • Per-image failures are logged while the remaining batch continues.

08

Technical challenges

  • Responsive desktop UI: Long-running file and API operations run in a background thread while a queue safely updates the Tkinter interface.
  • Credential handling: The application can store an API key in macOS Keychain without placing the secret in source files or command arguments.
  • Metadata compatibility: Values are mapped across modern XMP and legacy IPTC fields, including handling the legacy writer-field byte limit.
  • Model reliability: Prompts prohibit unsupported identities, events, relationships, and locations, while outputs remain drafts subject to human review.

09

Limitations

  • The desktop interface and Keychain integration are currently oriented toward macOS.
  • ExifTool and Python dependencies must be installed before processing files.
  • Vision-generated descriptions can still omit context or make mistakes and require human review.
  • Controlled-vocabulary matching is exact and does not yet model synonyms or hierarchical relationships.
  • The public source excludes production templates, operational records, credentials, and internal vocabularies.

10

Future improvements

  • Package the desktop application with a guided dependency installer.
  • Add cross-platform secure credential storage.
  • Create automated tests for intake parsing, field mapping, and metadata verification.
  • Add synonym-aware vocabulary review without allowing silent keyword approval.
  • Provide before-and-after metadata comparison inside the interface.

11

Code, demo, and supporting evidence

Open synthetic demoDownload desktop packagePython sourceSetup notes

The downloadable application is a public-safe copy. It preserves the architecture while replacing organization-specific configuration with generic templates and synthetic data. After installing the documented dependencies, macOS users can double-click the included launcher. The original operational file was not modified.

Related evolution: Metadata & Keyword Optimizer, the earlier prototype that explored controlled taxonomy and AI-assisted suggestions before this desktop workflow was developed.