Case study
How a K-12 Edtech Company Built a Personalized Learning Platform for Millions of Students
Empowering students with a platform that translates complex data into meaningful, personalized learning journeys.
One of the largest K-12 edtech companies in the US needed a personalized learning platform that could power products for millions of students. A prior vendor spent more than a year on the problem, only to create an AI tutor that hallucinated in front of the CEO. phData rebuilt it in 10 weeks, then stayed on to enhance the company’s entire AI product and R&D function. The platform now saves over $1M annually and delivers new AI use cases 5× faster than building without it.
annual OPEX savings from AI-powered curriculum automation
faster time-to-market for new AI use cases
At a glance
| Industry | K-12 Education Technology |
| Scale | One of the largest edtech companies in the US by revenue |
| Challenge | A prior AI vendor failed after $1.5M and 12+ months. The company needed a working personalized learning platform in 10 weeks. |
| Technology | Claude on Amazon Web Services Bedrock, Snowflake, AWS EKS, Redis, dbt, Neo4j |
| Timeline | 10-week initial proof of concept; ongoing multi-year AI product and R&D engagement |
| Result | $1M+ annual OPEX savings; 5× faster time-to-market for new AI use cases; production AI tutor, curriculum factory, and personalized learning journeys |
| Phdata Service | AI and Machine Learning, Anthropic, AWS, Snowflake |
The problem
Overcoming edtech’s AI infrastructure hurdles
AI is disrupting K-12 education fast. The client, one of the largest edtech businesses in the US by revenue, was making a deliberate bet: become a product company, not just a services business. That meant building a data-centric product organization where most enterprises only have IT and business lines. It also meant needing AI infrastructure that didn’t exist yet. A prior vendor had spent $1.5M and 12 months trying to build it, then their AI agent hallucinated in front of the CEO. phData had 10 weeks to build something that worked.
The client’s data existed in raw, unorganized form across disparate systems, with no feature store, no vector infrastructure, and no agentic orchestration layer. The AI and education markets were both moving rapidly, requiring a platform flexible enough to absorb product pivots without rebuilding from scratch. And the stakes were high. The platform needed to power multiple product lines serving real students, and a second hallucination failure was not an option.
What phData did
Four decisions shaped the delivery.
01
phData built a shared intelligence platform before building any individual product
Rather than delivering a single AI tutor and moving on, phData established the foundational infrastructure first: Snowflake as the organized data layer, AWS as the AI application environment, a custom agentic orchestration system, a context management layer for reading and writing structured data in and out of Snowflake, and a rigorous evaluation and guardrail harness. The alternative was faster initial delivery of a single tool, but exponentially more rework per use case thereafter. This upfront investment is why the platform now delivers new AI products 5× faster than building without it.
02
Snowflake was chosen as the data foundation; AWS as the AI application layer
Data without context cannot power a product. The goal was not just clean data — it was coherent data, where every metric has one definition, always, and the meaning travels with the data into the model. The company’s raw curriculum, student performance, and state standards data needed to be organized and made AI-readable before any model could use it reliably. phData used Snowflake to structure and govern that data using feature store-like patterns, enabling the AI system to read and write context with integrity. AWS hosted the AI applications that consumed it. This separation of concerns meant each layer could evolve independently, and Claude could access precisely structured context rather than raw, unreliable inputs.
03
phData built a custom evaluation and guardrail harness before any production deployment
Instead of deploying and monitoring, phData built synthetic student simulations and behavior tests that every AI system had to pass before touching a real student. This was especially important given the prior vendor’s hallucination failure and the K-12 compliance environment. It also made model selection rigorous and evidence-based rather than assumed: the platform could run any use case through the full test suite to determine which Claude model delivered the right performance at acceptable cost and latency.
04
Anthropic’s Claude won on the merits
phData’s eval suite established clear performance envelopes for each use case before any model was chosen. The Claude family consistently won those tests, with different tiers matched to different task requirements: Claude Haiku for fast, lightweight interactions and more capable Claude models for nuanced reasoning tasks such as tutoring, curriculum alignment, and multi-step content generation. For a K-12 product where a wrong answer has real consequences, that combination of reasoning quality, safety behavior, and tiered flexibility is what made Claude the right foundation for the entire platform.
The result
Scaling personalized learning with reliable AI infrastructure
The platform now powers multiple production AI products, saves over $1M annually, and delivers new use cases 5× faster.
Products include:
A K-12 AI tutor built on a Socratic approach, tested against synthetic students before any real deployment.
A context-first curriculum system that uses Claude to create thousands of editable components in under a minute at 90%+ accuracy, replacing an $80K–$150K manual process.
A personalized learning experience that tracks student performance and surfaces what to study next.
Tutoring session summaries trained on domain-specific patterns, more accurate than generic transcription tools.
A multi-modal content factory that ingests legacy materials and outputs problems, solutions, alt text, and grade-level variants at scale.
phData now enhances the company’s entire AI product and R&D function. The engagement has continued for multiple years, absorbing significant product and organizational change, because the platform was built to be sustainable and fast, not just impressive on a single demo.
annual OPEX savings from AI-powered curriculum automation
faster time-to-market for new AI use cases
Ready to see what a similar engagement could look like for your organization?
What the client said
Just because you have Snowflake or Claude doesn't mean you know how to do it. phData built what no one else could.
Why this matters beyond this project
Building a personalized learning platform that works in production requires more than infrastructure. It requires data that is coherent, not just clean, where context is preserved, domains are defined, and every metric means one thing, always. Most organizations skip this step. They build the orchestration layer, pick a model, and wonder why accuracy degrades. The frame has to come before the generation. phData builds both.
Frequently asked questions
What is a personalized learning platform and how does AI make it work?
A personalized learning platform tailors content, pacing, and feedback to individual students using AI. Building one that works in production requires structured data infrastructure, agentic orchestration, and rigorous model evaluation.
How long does it take to build a production-ready personalized learning platform?
phData delivered a working proof of concept in 10 weeks, then matured the platform over months. Timeline depends on data readiness and organizational alignment. Building the foundational infrastructure first means each subsequent use case ships faster.
Why did you choose Anthropic’s Claude over other AI models for this edtech platform?
Every use case ran through a custom eval suite before a model was chosen. The Claude family won consistently across tutoring accuracy, content generation quality, and safe behavior in K-12 contexts.
How do you prevent AI hallucinations in an educational product serving real students?
phData built a custom eval and guardrail harness that tests every AI system against synthetic student simulations before production deployment. This catches failure modes early and confirms safe behavior at scale. It is how a hallucinating PoC became a trusted daily tutor.
What technologies does phData use to build AI platforms for edtech companies?
Claude on AWS Bedrock, Snowflake, AWS EKS, Redis, dbt, and Neo4j. phData also built a custom agentic orchestration system and evaluation harness tailored to the client’s K-12 requirements.
Related case studies
Technology
AI & Data Strategy for a National Procurement Platform
Uncover how governed metrics, master data, and contract intelligence powered a 24‑month AI & Data plan for a national procurement platform.
Technology
How Order.co Used Agentic AI in Procurement to Automate Vendor Ordering
See how phData used agentic AI in procurement to automate Order.co’s vendor ordering in 6 weeks, with a 100% success rate using Anthropic Claude on AWS.
Technology
Biotech Leader Transforms and Scales Analytics Platform on AWS
Explore how a global biotech leader worked with phData to modernize analytics on AWS—empowering 2,000+ users with faster, more reliable insights across the org.
Get started
Ready to scope your own AI-powered personalized learning platform on Snowflake?
Talk to a phData expert about translating your data infrastructure into a production-ready AI platform. We’ll help you assess scope, timeline, and the right Snowflake services for your stack.
- Free scoping conversation — no commitment
- Snowflake-certified data and AI engineers
- Proven delivery
