Modernizing EdTech AI at Scale: How Jedi Hill Is Helping Xenontech Move from Azure/OpenAI to AWS Bedrock
Xenontech, a K-12 education technology platform serving schools across New South Wales, is preparing to move its AI infrastructure from Microsoft Azure and OpenAI to AWS Bedrock — with Jedi Hill designing and delivering the migration.
The Challenge
Xenontech's current AI capabilities run on Microsoft Azure infrastructure (VMs, Cosmos DB, Blob Storage) calling OpenAI's GPT-4 API. As usage has grown, three structural limitations have become clear:
No tiered model routing. Every request — from simple classification queries (roughly 74% of total traffic) to complex reasoning tasks — is processed through the same model and cost tier. At current volume, this single-model approach is projected to cost approximately $145,000/month.
Inefficient content safety architecture. Azure Content Safety requires a separate post-inference API call, adding latency and cost to every request.
Context window constraints. GPT-4o's 128K token limit means Xenontech's curriculum content library can't be ingested in a single pass, requiring additional chunking logic and engineering overhead.
The Planned Solution
Jedi Hill is designing a migration to a VPC-isolated, serverless-first architecture on AWS in the ap-southeast-2 (Sydney) region, with Amazon Bedrock accessed exclusively through a VPC PrivateLink Interface Endpoint — ensuring inference traffic never traverses the public internet.
Target architecture (five layers):
Edge & Access — Amazon CloudFront, AWS WAF, and Application Load Balancer handle TLS termination and rate limiting; Amazon Cognito manages authentication across separate student, teacher, and B2B API user pools.
Application — Amazon API Gateway routes requests through AWS Lambda, with Amazon SQS and AWS Step Functions plus EC2 Auto Scaling handling orchestration and asynchronous workflows.
AI/ML — Amazon Bedrock (accessed via PrivateLink) with a tiered routing strategy: Claude Haiku 4.5 handles high-volume intent classification, Claude Sonnet 4.6 handles RAG queries, code generation, and long-context tasks, and Claude Opus handles complex reasoning for a small fraction of critical-event traffic. Amazon Bedrock Guardrails screens both inputs and outputs.
Data — Amazon DynamoDB, Amazon S3, Amazon OpenSearch Serverless (vector knowledge base using Titan Text Embed v2), and Amazon ElastiCache for response caching.
Observability — Amazon CloudWatch, AWS Cost Explorer, and AWS Budgets provide per-model cost attribution and budget alerting.
Figure: Target AWS architecture for the Xenontech migration — from CloudFront/WAF edge access through Lambda-based application logic to a PrivateLink-isolated Amazon Bedrock inference layer, backed by DynamoDB, OpenSearch, and full observability/FinOps tooling.
Why Claude on Bedrock over continuing with GPT-4: the evaluation compared seven models/options across cost, context window, latency, and safety architecture. Claude's 200K context window fully accommodates Xenontech's curriculum corpus in a single pass (versus GPT-4o's 128K limit); Constitutional AI training embeds safety reasoning natively into inference rather than requiring a separate API call; and a single AWS account architecture eliminates cross-cloud IAM complexity.
Projected Impact
Based on modeling completed during the feasibility phase, the migration targets:
~37% reduction in AI inference cost through tiered model routing (from ~$145,000/month to a projected ~$92,000/month)
Elimination of separate content-safety API costs via Bedrock Guardrails and Constitutional AI
100% single-pass document processing (up from requiring chunking under the 128K limit)
35% of Tier-1 support queries resolved autonomously by an AI agent (up from ~0% today)
~50% reduction in manual scripting workload for the backend team through AI-assisted script generation
About Jedi Hill
Jedi Hill is an AWS Partner specializing in generative AI consulting and cloud-native application delivery, helping enterprises design and deliver production-grade AI platforms on AWS.
Interested in modernizing your AI infrastructure on AWS? Contact Jedi Hill to learn more.
Comments