AI Product Architecture
A Modular, API-First Learning Platform
This chapter documents the architecture behind the Yourclass platform's AI assessment, recommendation, and analytics capabilities: an API-first system built on AWS Amplify, Supabase, and the Claude API, with security and responsible-AI governance designed in from the start rather than added afterward.
- Context
- YourClass — Founder, Solution Architecture
- Period
- 2026 – Present
- Tools
- React, TypeScript, AWS Amplify, Supabase, Claude API
Business Challenge
An AI-enabled learning platform combining diagnostics, personality assessment, and synthesis reporting needed an architecture that was modular enough to evolve, secure enough to handle sensitive learner data, and simple enough for a small team to actually operate.
Architecture Approach
The platform is organized into four layers — client, application, data, and intelligence — each with a single clear responsibility and connected through defined API boundaries rather than direct cross-layer coupling.
- Client layer: a React frontend for diagnostics, assessments, and reports, hosted on AWS Amplify
- Application layer: routing and server-side session handling, keeping business logic off the client
- Data layer: Supabase (PostgreSQL) with Row-Level Security enforced on every table
- Intelligence layer: the Claude API powering the assessment engine, recommendations, and analytics
Outcome
- A modular intelligent learning platform architecture that can evolve layer by layer without a full rebuild
- Security-by-design: the Anthropic API key never leaves server-side secrets, and answer keys are stripped before any client response
- Row-Level Security enforced on every table, so data access is governed at the database layer, not just the application layer
- A governance boundary ensuring personality assessment scores never feed lead scoring or any commercial use beyond the learner's own report
Exhibit 14.1
Layered AI product architecture with security and governance controls
Lessons learned
- Deciding early that personality data would never touch lead scoring wasn't just an ethical choice. It also simplified the architecture, because that boundary didn't have to be re-litigated in every later feature discussion.
- Row-Level Security at the database layer is worth the upfront design cost; it means a future application-layer bug can't accidentally expose another learner's data.
Consultant's note
"The decision to strip answer keys server-side before any response reaches the client sounds like a small implementation detail. It's really a governance decision disguised as an engineering one, and it's much easier to build correctly on day one than to retrofit once the platform has real users."
Skills demonstrated
- AI Product Architecture
- API Design
- Data Governance
- Responsible AI
Part of Volume 3 — Delivery Excellence, Product Strategy & Enterprise Transformation.