Making EPUBs Accessible with AI: A Serverless Pipeline for Automated Alt Text Generation
summary
A serverless pipeline generates alt text for EPUB images at scale, handling ~130,000 images from 160 books per year, with editorial review in the loop and production release due in September. It extracts images with ebooklib + BeautifulSoup and prompts large multimodal models for captions, runs the API serverless via FastAPI + Mangum while dispatching heavy background jobs to AWS Fargate, stores metadata in a shared database, and uses Daxter for daily reliability checks.
AI-written summary · about
chapters (10)
- 0:00Introduction to alt text
- 2:11Why AI for alt text
- 3:58Monolithic solution limits
- 8:13Modular architecture split
- 11:49Docker optimizations
- 13:57Serverless deployment
- 17:33Reliability pipeline
- 20:25Limitations of alt text
- 23:15Advanced image classification
- 24:51Q&A session
related talks
48mBuilding Serverless AI powered Applications on AWS
28mAutomation Engineering with Serverless Compute
33mYour Next Publishing Product Should Be Aimed at AI Agents
01h 00mAI and alt text
44mAutomatically scale a serverless app with Amazon Textract & MongoDB
38mBuilding Practical, Cost-Efficient GenAI Solutions Using Serverless
57mUsing AI to automate content moderation and compliance with 
50mDeveloping Highly Scalable Image Storage Solution with AWS Serverless