Content Accessibility Utility on AWS

Digital content stakeholders across industries aim to streamline how they meet accessibility compliance standards efficiently. The “Content Accessibility Utility on AWS” offers a comprehensive solution for modernizing web content accessibility with state-of-the-art Generative AI models, powered by Amazon Bedrock. “Content Accessibility Utility on AWS” allows users to automatically audit and remediate WCAG 2.1 accessibility compliance issues. To get started, the solution offers a Python CLI and API. Capabilities currently include batch processing capabilities for handling large volumes of content efficiently, usage tracking to enable detailed cost management, and will continue to expand capabilities to support other content type and modals.

Table of Contents

• Features
• Prerequisites
• Installation
• Configuration
• Architecture
• Core Packages
  • PDF2HTML
  • Audit
  • Remediate
  • Batch
• Command Line Interface
• Python API
• Requirements
• License

Features

• Convert PDF documents to accessible HTML
• Preserve layout and visual appearance
• Extract and embed images
• Audit HTML for WCAG 2.1 and 2.2 accessibility compliance
• Remediate common accessibility issues using Bedrock models
• Advanced table remediation strategies
• Optional browser-backed (rendered) audit that detects computed-style and interactive issues static HTML analysis cannot see (e.g. focus visibility), using a real headless browser and axe-core
• Optional accessibility agent (Strands) that drives a render → fix → verify loop, confirming each fix actually renders correctly before marking it resolved
• Optional internationalization ([i18n]) that translates the worked-on content into one or more target languages via Amazon Bedrock — preserving markup and screen-reader-announced attributes (alt, title, aria-label), setting per-language lang/dir (WCAG Language of Parts), and optionally emitting a single multilingual document with an accessible language selector and browser-language auto-detection
• Support for single-page and multi-page output formats
• Batch processing capabilities for large-scale document processing
• Detailed usage tracking for BDA pages and Bedrock tokens
• Cost analysis tools for resource usage monitoring
• Streamlit sample web interface with usage visualization

Prerequisites

Before using the Content Accessibility with AWS tool, ensure the following prerequisites are met:

1. AWS Account: You need an AWS account with appropriate permissions.

2. S3 Bucket: Create an S3 bucket for storing input files, intermediate results, and outputs.

   aws s3 mb s3://my-accessibility-bucket
   

3. BDA Project: Set up an AWS Bedrock Data Automation (BDA) project.

   aws bedrock-data-automation create-data-automation-project \
       --project-name my-accessibility-project \
       --standard-output-configuration '{"document": {"extraction": {"granularity": {"types": ["DOCUMENT", "PAGE", "ELEMENT"]},"boundingBox": {"state": "ENABLED"}},"generativeField": {"state": "DISABLED"},"outputFormat": {"textFormat": {"types": ["HTML"]},"additionalFileFormat": {"state": "ENABLED"}}}}'
   

   Note the projectArn from the output, as it will be required for processing.

4. AWS CLI Configuration: Configure AWS credentials and default region.

   aws configure
   

Installation

# From PyPI
pip install content-accessibility-utility-on-aws

# From source
pip install .

Optional extras

The core install is static-only (no browser). The browser-backed audit and the accessibility agent are opt-in extras so the base dependency footprint stays small:

# Rendered audit: detect computed-style / interactive issues in a real browser
pip install "content-accessibility-utility-on-aws[rendered]"

# Agent: the render -> fix -> verify loop (implies the rendered layer)
pip install "content-accessibility-utility-on-aws[agent]"

# i18n: translate content into target languages + multilingual output
pip install "content-accessibility-utility-on-aws[i18n]"

Both extras use a headless browser via Playwright. After installing either extra, download the browser binary once:

playwright install chromium

The core package never imports the browser or agent stack, so existing static-only workflows are unaffected. When an extra is not installed, the --rendered/--agent options log a warning and fall back to the static audit. For running the rendered layer in AWS without bundling a browser, see the Rendered & Agent Guide (Amazon Bedrock AgentCore Browser Tool).

Deploy the managed pipeline (pip only, no repo checkout)

A complete event-driven pipeline — upload a document to S3 → convert (PDF) → audit → agent-remediate → accessible result written back to S3 — ships with the package. Scaffold the deployment files and deploy them; no clone required:

pip install "content-accessibility-utility-on-aws[agent]"
pip install bedrock-agentcore-starter-toolkit aws-sam-cli

# One interactive command scaffolds the files and runs the whole deploy —
# agentcore configure -> launch -> sam deploy — prompting for region, bucket,
# and (for the PDF path) BDA config, and wiring the runtime ARN between steps
# for you. Each cloud step is confirmed; add --yes for CI, --dry-run to preview.
content-accessibility-utility-on-aws deploy-pipeline

# Write the SAM template, AgentCore runtime app, and trigger Lambda into a dir:
content-accessibility-utility-on-aws init-pipeline ./a11y-pipeline
cd a11y-pipeline

agentcore configure --entrypoint agentcore_app.py --name a11y_pipeline \
  --requirements-file requirements.txt --region <region>
# For the PDF path, pass BDA config as runtime env vars; note the runtime ARN.
agentcore launch --env BDA_S3_BUCKET=<bucket> --env BDA_PROJECT_ARN=<bda-project-arn>
sam deploy --guided --parameter-overrides \
  AgentRuntimeArn=<runtime-arn> InputBucketName=<globally-unique-bucket>

Then upload documents to the input bucket (pdf/ for PDFs, html/ for HTML or a .zip of HTML+CSS+JS); results land under accessible/. The generated README.md and the Rendered & Agent Guide cover IAM, BDA setup, and usage in full.

Configuration

Environment Variables

Set the following environment variables to configure the tool:

export BDA_S3_BUCKET=my-accessibility-bucket
export BDA_PROJECT_ARN=arn:aws:bedrock:us-west-2:123456789012:project/my-accessibility-project

Optional environment variables:

• AWS_PROFILE: Specify an AWS CLI profile to use.
• CONTENT_ACCESSIBILITY_WORK_DIR: Directory for temporary files (default: system temp).

Example Configuration File

The tool supports configuration files for easier setup. Below is an example configuration file (my-config.yaml):

# PDF conversion settings
pdf:
  extract_images: true
  image_format: png
  embed_images: false
  single_file: true
  continuous: true
  embed_fonts: false
  exclude_images: false
  cleanup_bda_output: false

# Accessibility audit settings
audit:
  audit_accessibility: true
  min_severity: minor
  detailed_context: true
  skip_automated_checks: false
  issue_types: null  # Set to a list of specific issue types or null for all

# Remediation settings
remediate:
  max_issues: 100
  model_id: us.anthropic.claude-sonnet-5
  issue_types: null
  severity_threshold: minor
  report_format: json

# AWS settings
aws:
  # To use an existing BDA project:
  create_bda_project: false
  bda_project_arn: "arn:aws:bedrock:us-west-2:123456789012:project/my-accessibility-project"
  
  # OR to create a new BDA project:
  # create_bda_project: true
  # bda_project_name: "my-new-accessibility-project"
  
  s3_bucket: my-accessibility-bucket

Architecture

The package consists of four main modules working together to convert, audit, remediate, and batch process documents, plus an optional browser-backed agent layer on top:

graph TD
    A[PDF2HTML] --> B[Convert PDFs to HTML]
    A --> C[Extract & Process Images]
    D[Audit] --> E[Check Accessibility Issues]
    F[Remediate] --> G[Fix Accessibility Problems]
    F --> H[Generate Remediation Reports]
    I[Batch] --> J[Orchestrate Large-scale Processing]
    I --> K[Track Jobs & Handle AWS Integration]
    L[Agent - optional] --> M[Render in a real browser]
    L --> N[Detect computed-style & interactive issues]
    L --> O[Apply fix, re-render, verify]

    A --> I
    D --> I
    F --> I
    D --> L
    F --> L

The optional agent layer (agent/) renders pages in a real headless browser to find issues static analysis cannot (computed contrast, focus visibility, the accessibility tree) and closes the loop by re-rendering to verify each fix. It is fully additive and off by default. See the Rendered & Agent Guide.

Core Packages

PDF2HTML

The PDF2HTML module handles conversion of PDF documents to HTML, including image extraction and processing.

graph TD
    A[PDF Source] --> B[PDF2HTML]
    B --> C[BDA Integration]
    B --> D[Image Processing]
    B --> E[HTML Generation]
    C --> F[HTML Output]
    D --> F
    E --> F

Key components:

• Bedrock Data Automation (BDA) integration for PDF parsing
• Image extraction and processing
• HTML structure generation with preserved layout
• Support for both single-page and multi-page output

Audit

The Audit module analyzes HTML for accessibility issues according to WCAG 2.1 and 2.2 guidelines.

graph TD
    A[HTML Input] --> B[Audit Module]
    B --> C[Document Checks]
    B --> D[Structure Checks]
    B --> E[Image Checks]
    B --> F[Table Checks]
    C --> G[Audit Report]
    D --> G
    E --> G
    F --> G

Key components:

• Comprehensive accessibility checks
• Issue severity classification
• Detailed context information
• Multiple report formats (HTML, JSON, text)

WCAG 2.2 scope: Because this tool produces static HTML converted from PDFs, it audits and remediates the WCAG 2.2 criteria that apply to non-interactive documents — currently 2.5.8 Target Size (Minimum). The remaining 2.2 criteria (2.5.7 Dragging Movements, 3.2.6 Consistent Help, 3.3.7 Redundant Entry, 3.3.8 Accessible Authentication) govern interactive web-application behaviors such as drag gestures, multi-page help, and authentication flows, which are out of scope for generated document content.

Remediate

The Remediate module fixes accessibility issues identified during audit.

graph TD
    A[HTML with Issues] --> B[Remediate Module]
    B --> C[AI Remediation Strategies]
    B --> D[Direct Fixes]
    C --> E[Remediated HTML]
    D --> E
    B --> F[Table Remediation]
    F --> G[Direct Table Fixes]
    F --> H[AI-Powered Table Fixes]
    G --> E
    H --> E

Key components:

• AI-powered remediation using Bedrock models
• Direct fixes for common issues
• Advanced table structure remediation
• Image accessibility enhancements
• Remediation reporting

Batch

The Batch module provides orchestration for processing documents at scale.

graph TD
    A[Document Source] --> B[Batch Module]
    B --> C[Job Management]
    B --> D[AWS Integration]
    B --> E[Processing Pipeline]
    C --> F[Status Tracking]
    D --> G[S3 & DynamoDB]
    E --> H[Lambda Integration]
    F --> I[Job Completion]
    G --> I
    H --> I

Key components:

• AWS service integrations
• Job tracking and status management
• Asynchronous processing
• Lambda function support

Command Line Interface

The package provides a command-line interface with several subcommands:

PDF to HTML Conversion

content-accessibility-utility-on-aws convert --input path/to/document.pdf --output output/directory

Options:

• --single-file: Generate a single output file
• --single-page: Combine all pages into a single HTML document
• --multi-page: Keep pages as separate HTML files
• --extract-images: Extract and include images from the PDF (default: True)
• --image-format [png|jpg|webp]: Format for extracted images
• --embed-images: Embed images as data URIs in HTML
• --s3-bucket: Name of an existing S3 bucket to use
• --bda-project-arn: ARN of an existing BDA project to use
• --create-bda-project: Create a new BDA project if needed
• --config: Path to configuration file

Accessibility Audit

content-accessibility-utility-on-aws audit --input path/to/document.html --output accessibility-report.json --format json

For HTML report:

content-accessibility-utility-on-aws audit --input path/to/document.html --output accessibility-report.html --format html

To additionally run the browser-backed (rendered) audit, which detects computed-style and interactive issues (e.g. focus visibility) the static audit cannot see:

content-accessibility-utility-on-aws audit --input path/to/document.html --output report.json --rendered

Options:

• --format, -f [json|html|text]: Output format for audit report
• --checks: Comma-separated list of checks to run
• --severity [minor|major|critical]: Minimum severity level to include in report
• --detailed: Include detailed context information in report (default: True)
• --summary-only: Only include summary information in report
• --rendered: Also render each page in a headless browser to detect computed-style/interactive issues static analysis misses (requires the [rendered] extra and playwright install chromium)
• --agent: Use the browser-backed agent for the rendered pass (implies --rendered; requires the [agent] extra)
• --config: Path to configuration file

Remediation

content-accessibility-utility-on-aws remediate --input path/to/document.html --output remediated.html

Options:

• --auto-fix: Automatically fix issues where possible
• --max-issues: Maximum number of issues to remediate
• --model-id: Bedrock model ID to use for remediation
• --severity-threshold [minor|major|critical]: Minimum severity level to remediate
• --audit-report: Path to audit report JSON file to use for remediation
• --single-page: Combine all pages into a single HTML document
• --multi-page: Keep pages as separate HTML files
• --generate-report: Generate a remediation report after remediation (default: True)
• --report-format [html|json|text]: Format for the remediation report
• --config: Path to configuration file

Translation (i18n)

Translate the worked-on HTML into one or more target languages. Requires the [i18n] extra. The translation itself runs on Amazon Bedrock (the same core dependency remediation uses); the extra adds localized language names and source-language auto-detection.

# One accessible HTML file per language (doc.es.html, doc.fr.html, ...)
content-accessibility-utility-on-aws translate \
  --input remediated.html --output out/ --target-languages es,fr,ja

# A single multilingual document with a language selector that auto-selects the
# visitor's browser language on first load
content-accessibility-utility-on-aws translate \
  --input remediated.html --output multilingual.html \
  --target-languages es,fr,ar --multilingual

Options:

• --target-languages (alias --languages): Comma-separated BCP-47 codes to translate into, e.g. es,fr,ja (required)
• --source-language: Source language; auto-detected from the document if omitted
• --multilingual: Emit one combined document with an accessible language selector instead of one file per language
• --no-language-selector: Omit the visible selector from multilingual output
• --no-browser-language: Do not auto-select the visitor's browser language on first load
• --model-id: Bedrock model ID to use for translation
• --config: Path to configuration file

The translator preserves all markup and translates screen-reader-announced attributes (alt, title, aria-label), sets lang/dir on each language block (WCAG 3.1.2 Language of Parts), and skips <script>/<style>/<code> and any element marked translate="no". The language selector and browser detection are progressive enhancements — with JavaScript disabled the default language stays visible and no content is hidden.

Complete Processing

content-accessibility-utility-on-aws process --input path/to/document.pdf --output output/directory

This command runs the full workflow:

1. Converts PDF to HTML
2. Audits the HTML for accessibility issues
3. Remediates the issues found
4. Optionally translates the remediated HTML (when --target-languages is given)

Options:

• --skip-audit: Skip the audit step
• --skip-remediation: Skip the remediation step
• --audit-format [json|html|text]: Format for the audit report
• --severity [minor|major|critical]: Minimum severity level for audit and remediation
• --auto-fix: Automatically fix issues where possible
• --rendered: Include the browser-backed rendered audit (see Audit above)
• --agent: Use the browser-backed agent for the rendered pass (implies --rendered)
• --target-languages es,fr,...: Also translate the result (requires [i18n]; see Translation above). Add --multilingual for a single selectable document. Per-language files are written to a translations/ subdirectory.
• Plus all options available in the individual commands
• --config: Path to configuration file

Scaffold the managed cloud pipeline

content-accessibility-utility-on-aws init-pipeline ./a11y-pipeline

Writes the deployment files (SAM template, AgentCore runtime app, trigger Lambda, requirements) into the given directory so you can deploy the event-driven S3 pipeline without checking out the repository. See Deploy the managed pipeline.

Options:

• --force: Overwrite existing files in the target directory

Deploy the managed cloud pipeline (interactive)

content-accessibility-utility-on-aws deploy-pipeline

Scaffolds the files and runs the whole deploy — agentcore configure → agentcore launch → sam deploy — prompting for values and wiring the runtime ARN between steps. Requires the agentcore and sam CLIs on PATH.

Options:

• --region, --input-bucket, --bda-bucket, --bda-project-arn: set values non-interactively (anything omitted is prompted for)
• --runtime-name: AgentCore runtime name (default a11y_pipeline)
• --yes / -y: unattended (CI) — skips the per-step confirmations and runs sam deploy non-interactively (explicit flags instead of --guided)
• --dry-run: print the exact commands and exit without running anything
• --force: overwrite existing scaffold files

Use a configuration file

content-accessibility-utility-on-aws convert --config my-config.yaml --input document.pdf

Override config file settings with command-line arguments

content-accessibility-utility-on-aws audit --config my-config.yaml --severity major --input document.html

Common Options

These options are available for all commands:

• --input, -i: Input file or directory path (required)
• --output, -o: Output file or directory path (defaults to a path based on input name)
• --debug: Enable debug logging
• --quiet, -q: Only output reports, suppress other output
• --config, -c: Path to configuration file
• --profile: AWS profile name to use for credentials

Output Structure

Convert Command Output

output-directory/
├── extracted_html/              # Directory with HTML files
│   ├── document.html            # Combined HTML file (if --single-file)
│   ├── page-0.html              # Individual page files (if not --single-file)
│   ├── page-1.html
│   └── ...
└── images/                      # Directory with extracted images
    ├── image-0.png
    ├── image-1.png
    └── ...

Process Command Output

output-directory/
├── html/                        # Converted HTML + extracted images
├── audit_report.[json|html|txt] # Audit report
└── remediated_<name>.html       # Final remediated HTML
                                  #   (remediated_document.html with --single-page;
                                  #    remediated_html/ with --multi-page)

Streamlit Sample Web Interface

A sample Streamlit web interface has been developed to demonstrate the functionality of the Document Accessibility tool. This interface allows users to upload documents, configure processing options, and view results interactively. To learn more about the Streamlit interface, refer to the Streamlit Guide.

Python API

The package provides a Python API for programmatic use:

Complete Processing Pipeline

from content_accessibility_utility_on_aws.api import process_pdf_accessibility

# Process a PDF through the full pipeline
result = process_pdf_accessibility(
    pdf_path="document.pdf",
    output_dir="output/",
    conversion_options={
        "single_file": True,
        "image_format": "png"
    },
    audit_options={
        "severity_threshold": "minor",
        "detailed": True
    },
    remediation_options={
        "model_id": "us.anthropic.claude-sonnet-5",
        "auto_fix": True
    },
    perform_audit=True,
    perform_remediation=True
)

Individual Components

from content_accessibility_utility_on_aws.api import (
    convert_pdf_to_html,
    audit_html_accessibility,
    remediate_html_accessibility
)

# Convert PDF to HTML
conversion_result = convert_pdf_to_html(
    pdf_path="document.pdf",
    output_dir="output/",
    options={
        "single_file": True,
        "image_format": "png"
    }
)

# Audit HTML for accessibility issues
audit_result = audit_html_accessibility(
    html_path="output/document.html",
    options={
        "severity_threshold": "minor",
        "detailed_context": True
    }
)

# Remediate accessibility issues
remediation_result = remediate_html_accessibility(
    html_path="output/document.html",
    audit_report=audit_result,
    options={
        "model_id": "us.anthropic.claude-sonnet-5",
        "auto_fix": True
    }
)

Translation (i18n)

Translate the (remediated) HTML into one or more target languages. Requires the [i18n] extra. Returns the source language, the target languages, and a mapping of language → output file path.

from content_accessibility_utility_on_aws.api import translate_html_accessibility

# One accessible file per language
result = translate_html_accessibility(
    html_path="output/remediated.html",
    options={"target_languages": ["es", "fr", "ja"]},
    output_path="output/translations/",
)

# Or a single multilingual document with a language selector that auto-detects
# the visitor's browser language on first load
result = translate_html_accessibility(
    html_path="output/remediated.html",
    options={
        "target_languages": ["es", "fr", "ar"],
        "multilingual": True,
        "add_language_selector": True,   # default
        "use_browser_language": True,    # default
    },
    output_path="output/multilingual.html",
)

Browser-backed (rendered) audit and agent

The rendered layer is enabled through the same audit_html_accessibility API by setting options["rendered"] (or options["agent"]). Rendered findings use the identical issue shape as the static audit, so the returned report and any downstream remediation work unchanged. Requires the [rendered]/[agent] extra and playwright install chromium.

from content_accessibility_utility_on_aws.api import audit_html_accessibility

# Static audit + rendered pass (adds e.g. focus-visible findings)
audit_result = audit_html_accessibility(
    html_path="output/document.html",
    options={"rendered": True},
    output_path="report.json",
)

To drive the full render → fix → verify loop directly with the Strands agent (returns the remediated HTML, the committed resolutions, and the agent's tool-call trace):

from content_accessibility_utility_on_aws.agent.browser_probe import make_browser_probe
from content_accessibility_utility_on_aws.agent.agent import run_agent

with open("output/document.html") as f:
    html = f.read()

# make_browser_probe() selects the browser backend from options/env:
#   local Playwright Chromium by default, or the managed AgentCore browser
#   when options["browser_backend"] == "agentcore" (see the guide below).
with make_browser_probe() as probe:
    result = run_agent(probe, html)

print(result["resolved"])   # issues confirmed fixed by a passing verify()
print(result["tool_log"])   # the agent's render/apply_fix/verify/commit trace

See the Rendered & Agent Guide for the architecture, the verify-before-commit guarantee, and cloud deployment on Amazon Bedrock AgentCore.

Batch Processing

The batch package is a set of per-stage processors plus S3/DynamoDB/SQS helpers, designed to be wired into an event-driven pipeline (e.g. Lambda functions triggered by S3 events or SQS messages). Each stage takes a job id and S3 locations, does its work, and writes results back to S3.

from content_accessibility_utility_on_aws.batch.common import (
    generate_job_id,
    create_job_record,
    get_job_status,
)
from content_accessibility_utility_on_aws.batch.pdf2html import process_pdf_document
from content_accessibility_utility_on_aws.batch.audit import process_html_document

# Create a job record (tracked in DynamoDB)
job_id = generate_job_id("my-bucket", "documents/file.pdf")
create_job_record(job_id, document_key="documents/file.pdf", stage="PDF_TO_HTML")

# Stage 1: convert a PDF from S3 to HTML, writing results back to S3
conversion = process_pdf_document(
    job_id=job_id,
    source_bucket="my-bucket",
    source_key="documents/file.pdf",
    destination_bucket="my-bucket",
    options={"single_file": True},
)

# Stage 2: audit the produced HTML (see batch.remediate for the remediation stage)
audit = process_html_document(
    job_id=job_id,
    source_bucket="my-bucket",
    source_key=conversion["html_key"],
    destination_bucket="my-bucket",
    options={"severity_threshold": "minor"},
)

# Inspect job status at any time
status = get_job_status(job_id)

In production these stages run as separate Lambda functions chained by S3/SQS events; batch.common provides parse_s3_event, parse_sqs_event, send_sqs_message, and update_job_status for that wiring.

Requirements

• Python 3.11+
• AWS credentials for Bedrock Data Automation and Bedrock models
• Appropriate IAM permissions for S3 and BDA services

For AWS credentials configuration:

1. Set up AWS CLI with aws configure
2. Use environment variables (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY)
3. Or specify a profile with the --profile option

License

Apache-2.0 License. See LICENSE for details.

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for details on how to contribute to this project.