AI-Driven Image-to-Alt-Text CI Automation

Overview

A transformer-powered image alt-text compliance system that automatically generates and validates accessible alt-text for images at CVS Health. The project began as a hackathon prototype and matured into a production CI pipeline.

Origin: 1st Prize at CVS GenAI Hackathon (Aug 2023)

Won 1st Prize for a proof-of-concept that used Hugging Face vision-language models to generate and compare alt-text for images and test them for accessibility compliance. Combined transformer-based caption generation with sentiment analysis to flag tone issues alongside coverage gaps.

Productionization

After the hackathon, established production-grade pipelines for enterprise rollout under the CPE web-core CI/CD standards. The system runs on every pull request and:

• Generates candidate alt-text for new images using vision-language transformers
• Compares against existing alt-text to flag missing or inaccurate copy
• Scores compliance against WCAG/ADA requirements with sentiment-aware checks
• Reports inline in GitHub Actions and Slack to keep accessibility shifted left

Impact

• 35% boost in accessibility conformance after hackathon prototype rollout (per A11yScore baseline)
• 15% reduction in downstream defects after productionization for enterprise rollout
• Catches visual/copy regressions traditional snapshot testing misses
• Foundation for the CVS accessibility org’s automated CI quality gates

Tech Stack

• AI/ML: Hugging Face Transformers, vision-language models, sentiment analysis
• CI/CD: GitHub Actions, custom pipeline orchestration, Slack reporting
• Backend: Python
• Validation: WCAG/ADA rule evaluation, A11yScore baseline metrics