Building Accessible AI-Powered Learning for a US K–12 Education Company

A client is a US-based education technology company operating across K–12 learning provides an LMS and interactive digital science materials for students and teachers.

The client needed help improving its AI-powered experiment generation pipeline, meeting accessibility requirements, and building AI tools for content quality and alignment with educational standards.

Key achievements

2.5×	faster task completion after refactoring the experiment generation pipeline and improving its development workflow.
WCAG	compliance across interactive experiments, with screen reader captions, reduced motion and keyboard-only control
7	quality criteria, including standards compliance and coherence, now checked automatically in content that curriculum experts used to review line by line

Challenge

The client had an established digital learning platform with interactive science experiments, an AI-powered content generation pipeline, and a team of subject matter experts creating educational materials.

As the product grew, several technical limitations started affecting how quickly the client could create, update, test, and deliver learning content.

A large codebase was slowing development
The client handed over a generation pipeline with roughly 2.5 million lines of code, delivered mainly as archives without a Git workflow, DevOps setup, or reliable versioning. Its monolithic structure made changes risky and maintenance difficult. The client was sometimes generating the same experiment repeatedly to get a usable result, which slowed work on new content.
Content review could not scale manually
Subject matter experts were checking educational materials against large sets of state and federal standards, often line by line. As the amount of content grew and standards changed, this manual process became difficult to sustain and made large-scale content updates impractical.
Lesson adaptation required repeated expert work
The client also needed to shorten, restructure, and adapt existing lessons when curricula or teaching time changed. Experts had to decide what could be removed while preserving required knowledge, while also checking content for contradictions, missing explanations, and unclear wording.
Interactive experiments needed more precise behavior
The platform turns real science experiments into interactive Unity simulations. The existing physics and scene-generation capabilities had limitations that affected both scientific accuracy and the quality of the student experience. The client needed more control over objects, lighting, and experiment conditions while keeping the simulations suitable for tablets and Chromebooks.
Accessibility had to work across the product
The experiments needed to meet WCAG requirements for students with different accessibility needs. This affected captions, screen-reader descriptions, motion, and interaction methods. Keyboard control also required additional engineering because some students could have difficulty with precise mouse movements.

Solution

We took over the existing AI and Unity pipeline, reorganized its development process, and expanded the platform’s capabilities for interactive science learning.

Our work covered the experiment generation pipeline, accessibility, content processing, translation, testing, and educational content management.

1
Checking content against state standards
We built an AI system that checks educational materials against state and federal standards. A standards document runs to about 50 pages of broadly worded requirements.

The system splits course content into chunks and builds a content graph that maps how topics connect across lessons and chapters. A single chapter can’t cover a whole standard, but it still has to meet some smaller requirements, and the graph lets the system check both the course and its individual parts.

For each fragment that falls short of a standard, the system shows its location and suggests a revision. Experts accept or reject every suggestion, so existing lessons stay in place and only the flagged parts change.
2
Seven criteria for content quality
The system evaluates content against seven criteria, so experts see quality
- issues alongside standards gaps:
- compliance with standards
- completeness
- coherence
- clarity
- contradictions between parts of the material
- terms used before they’re explained
- overcomplicated wording
Before the project, experts tracked these checks by hand in Excel.
3
Adapting lessons with the content condenser
The condenser restructures existing lessons when a curriculum or the available teaching time changes. For example, it can turn two lessons that take five days into one three-day lesson. It keeps the most important content and runs the same quality checks on the result. If a shortened version no longer covers what the standards require, the teacher sees warnings and makes the final decision.

The condenser produces teacher and student versions of the materials at once, because the two audiences need different formats.
4
Refactoring the experiment generation pipeline
Our engineers refactored the 2.5-million-line pipeline so that a change in one part no longer breaks others. We moved the code into Git and set up a Git flow and DevOps processes. We also reworked the generation logic so an experiment description produces a usable scene on the first run.
5
AI-assisted testing
The AI goes straight to the step under review and checks how it works. Testers no longer click through a full investigation to reach one step. We moved the test pipeline to AI and automated tests for many scenarios, and each failed check comes with a specific fix for the developer.
6
Interactive science experiments
The Aristek experts expanded the physics and interaction capabilities of the existing simulations. The team added more precise controls for objects and experiment conditions, including adjustable lamp height and direction, so the simulations could represent different experimental conditions.

The visual layer was also improved with more detailed object textures and more realistic lighting. Schools use these simulations in place of physical labs they can’t equip, running them on devices students already have, such as tablets and Chromebooks. So the team made sure every visual and physics change works on these devices.
7
Accessibility
We built WCAG support into the experiments: step captions for screen readers, a reduced motion mode and full keyboard control.

The pipeline generates each caption with enough context for a student using a screen reader to follow what happens in the scene. Keyboard control took the most engineering work, because students interact with experiments by dragging objects.
8
Adding multilingual content support
Our team built translation support for the client’s main languages, together with terminology management. The system can use a client-defined glossary so the same educational term receives consistent translations throughout the material.

We also added text-to-speech support with different voices and expressive delivery for different languages.
9
Moving AI model infrastructure
We moved the models behind the client’s generation pipeline from Fireworks to AWS Bedrock. The pipeline used DeepSeek, hosted on Fireworks, a platform that runs AI models on its own hardware. The client chose Bedrock to meet its security requirements.

Bedrock runs models inside AWS infrastructure, so the client manages access to the models and the data sent to them through AWS security controls. AWS also states that Bedrock doesn’t use customer data to train models.

Screenshots

Project scope

We started with a few proofs of concept to show the company we could take over its engineering work. As more of that work moved to us from other vendors, the team grew from three engineers to a cross-functional team of about a dozen specialists.

The core stages included:

1. Discovery and assessment	Review of the existing AI and Unity pipeline, development setup, accessibility requirements, and educational content workflows to define the scope of work.
2. Pilot development	Delivery of a set of pilots covering interactive science experiments, AI-generated content, accessibility scenarios, and educational content analysis.
3. Technical restructuring	Refactoring of the inherited pipeline and setup of the development and delivery processes required for ongoing product work.
4. Product expansion	Extension of the platform with improved experiment capabilities, multilingual support, text-to-speech, content analysis, and lesson adaptation tools.
5. Expert validation	Testing the new capabilities against real educational materials and use cases, with feedback from the client’s subject matter experts guiding further adjustments.
6. Production rollout	Finalization of the WCAG-related work and deployment of the updated experiment experience and supporting capabilities to the client’s production environment.

Team

Project manager x1
Frontend engineer x3
Back-end Developers x2
Data Scientists x2
QA Engineers x3

Results

The first phase of the project gave the company a pipeline its engineers can change safely, experiments that more students can use, and a standards review process its experts can complete.

1
Faster development on the experiment pipeline
Development tasks on the pipeline now run about 2.5 times faster. The gain comes from the architecture our engineers designed during refactoring, with AI assigned well-defined tasks on top of it. Changes no longer trigger cascading failures across the codebase. Each new experiment generates in one run, down from around ten.
2
Experiments open to more students
After the WCAG release, students who use screen readers, need reduced motion or can’t use a mouse complete the same experiments as their classmates. In all six languages, terminology stays consistent from step to step, and each language gets its own voiceover.
3
Standards review the experts can complete
A manual review of the full content library used to be impossible to estimate. Now experts accept or reject the tool’s suggestions and revise only flagged fragments. The condenser adapts lessons to a new schedule in one step and replaces the experts’ Excel tracking.
4
Shorter testing cycles
AI-assisted and automated tests reduced the number of rounds between testers and developers. When a check fails, developers get a specific fix to make and can verify one step without replaying a 48-step investigation.

Next steps

The WCAG and experiment work represents the first part of a larger project. The team is continuing development of AI-powered tools for the client’s educational content workflows.

The next stage will include:

Teacher dashboards – dashboards showing how students are progressing and helping teachers identify students who may need different explanations or additional attention.
Adaptive learning paths – recommendations based on student learning data, helping teachers decide how content can be presented differently for individual students.
Standards management – further development of the standards-checking system, with plans to connect it directly to the client’s content management system and keep standards available within the workflow.

Together, these capabilities will give the client a way to review, adapt, and personalize educational content at a much larger scale while keeping subject matter experts responsible for the final content decisions.

Key insights

	AI-generated code still needs human architecture. A large volume of AI-generated code can create a system that works in isolation while becoming increasingly difficult to change as a whole. Clear architecture and task design remain essential.
	Accessibility can shape engineering decisions from the start. Requirements such as keyboard control, screen-reader descriptions, and reduced motion affect how interactions and content are designed, especially in highly interactive learning products.
	Educational content needs relationships, not document-level checks. Standards and learning objectives can span multiple chapters and lessons, so useful automated review needs to understand how pieces of content relate to each other.
	AI works best as a review layer when experts retain control. Automated checks can scan large volumes of material and point experts toward potential issues, while teachers and subject matter experts remain responsible for deciding what should change.

If your experts still check content against standards by hand, or your AI features have become hard to maintain, our team can review your setup and suggest where to start.

Building Accessible AI-Powered Learning for a US K–12 Education Company

Key achievements

Key achievements

Challenge

A large codebase was slowing development

Content review could not scale manually

Lesson adaptation required repeated expert work

Interactive experiments needed more precise behavior

Accessibility had to work across the product

Solution

Checking content against state standards

Seven criteria for content quality

Adapting lessons with the content condenser

Refactoring the experiment generation pipeline

AI-assisted testing

Interactive science experiments

Accessibility

Adding multilingual content support

Moving AI model infrastructure

Screenshots

Project scope

1. Discovery and assessment

2. Pilot development

3. Technical restructuring

4. Product expansion

5. Expert validation

6. Production rollout

Team

Project manager x1

Frontend engineer x3

Back-end Developers x2

Data Scientists x2

QA Engineers x3

Results

Faster development on the experiment pipeline

Experiments open to more students

Standards review the experts can complete

Shorter testing cycles

Next steps

Key insights

Key insights

If your experts still check content against standards by hand, or your AI features have become hard to maintain, our team can review your setup and suggest where to start.