Architecture & business story

ARCHITECTURE

One source of truth, end to end

Amazon Aurora PostgreSQL + pgvector stores cert objectives, uploaded PDF chunks, embeddings, generated questions, and every review event. Coverage, citations, and scheduling all read from the same data.

Amazon Bedrock powers retrieval and generation: Titan Text Embeddings V2 for semantic search over lecture pages, and Claude Sonnet for original MCQs tied to source passages.

FSRS schedules when each card is due based on recall history. BKT estimates per-objective mastery for dashboard readiness bars.

The app runs on Vercel — serverless API routes connect to Aurora over a secure connection.

SCHEMA

Core tables in Aurora PostgreSQL

• objectives — seeded CLF-C02 domains and tasks
• documents — uploaded PDF metadata and status
• chunks — page-level text with vector(512) embeddings
• items — generated MCQs with source page citations
• review_state — current adaptive review card state
• review_log — immutable answer and review events
• mastery — personalized mastery per objective
• ingestion_jobs — upload processing progress

CREATE EXTENSION IF NOT EXISTS vector; with HNSW index on chunk embeddings for fast similarity search.

VECTOR SEARCH

Example pgvector query

Retrieval finds the most relevant lecture pages before question generation:

SELECT c.content, c.page_number, d.filename
FROM chunks c
JOIN documents d ON d.id = c.document_id
ORDER BY c.embedding <=> $query_embedding
LIMIT 5;

BUSINESS MODEL

B2C today, B2B cohorts next

Free — one cert goal, limited uploads, cited practice questions, and adaptive review.

Pro (~$12–15/month) — multiple goals, unlimited PDFs, full dashboard readiness, and priority generation.

B2B cohort analytics— team readiness and instructor dashboards — on the roadmap for bootcamps and enterprise L&D.