Scaling AI-Personalized Learning to Production

A workforce development nonprofit trains underrepresented young adults in technology careers through structured apprenticeship programs. A consistent curriculum works at scale but leaves little room to address each apprentice's individual gaps. After an initial proof of concept demonstrated the feasibility of AI-driven lesson personalization, the organization committed to bringing that capability into production — a deliberate investment in long-term learning outcomes and a signal to funders that it is ready to lead on AI-powered program delivery.

Protagona's objective was to evolve the proof of concept into a production-ready personalized learning platform capable of generating 100 lesson plans per week. The system needed to incorporate real-time student skill-level data, surface the full lesson content library, and give instructors a structured interface to review, provide feedback on, and regenerate AI-produced lesson recommendations.

Moving from proof of concept to production introduced a different class of problems. The original build validated the core AI logic but was not designed to handle concurrent lesson generation at volume, integrate live student profile data, or support the instructor feedback loops required to build trust in AI output. Designing an architecture that could process asynchronous generation requests reliably — while supporting synchronous instructor interactions like review and regeneration — required careful orchestration across multiple AWS services. A nonprofit environment also demanded cost efficiency and low maintenance overhead, meaning every architectural decision had to favor managed services, disciplined scoping, and a clear MVP boundary.

Protagona built a serverless, event-driven architecture that separates asynchronous lesson generation from synchronous instructor-facing interactions. Generation requests are queued and processed independently, allowing the system to reach 100 lessons per week without blocking the instructor API or creating bottlenecks at peak usage. Amazon Bedrock powers the generative AI layer, drawing on each student's skill-level metadata and the full lesson catalog to produce differentiated recommendations across three tiers: foundational reinforcement, targeted practice, and extension work for high performers.

‍

An extended API layer gives instructors a single workflow to request a lesson plan, review AI output, annotate it, and trigger regeneration — keeping instructors in control of what reaches students. Production readiness was treated as a first-class requirement: structured logging, CloudWatch alarms, and health checks provide operational visibility without dedicated engineering oversight, and a CI/CD pipeline ensures future updates deploy consistently and safely.