Relearnable

Overview

Relearnable is a full-stack AI learning SaaS I designed and deployed to help users genuinely retain what they learn. The core innovation is a Knowledge Correction Engine—an LLM-driven system that identifies gaps in understanding and generates targeted curriculum in real time.

Problem

Most learning platforms measure completion, not comprehension. Users finish courses but forget the material weeks later. There's no feedback loop that catches misconceptions early or adapts content based on what a learner actually struggles with.

Constraints

• System needed to support live users at production scale
• LLM responses had to be fast enough for real-time interaction
• Mobile-first UX with minimal friction for repeat sessions
• Curriculum logic needed to be flexible across subjects

Solution

I built a Knowledge Correction Engine using RAG pipelines and vector embeddings to detect where learners have misconceptions. When gaps are identified, the system adaptively rebuilds understanding through targeted curriculum trees—structured learning paths that start precisely at the point of confusion. This approach improved user placement test scores by 40%, validated by educators who discovered gaps in their own expertise.

Architecture & Stack

Key Engineering Decisions

• RAG pipelines with vector embeddings for semantic misconception detection
• Hybrid caching strategy (Redis + PostgreSQL) cutting API costs by 30%
• Structured LLM outputs with schema validation for reliable curriculum generation
• PostgreSQL relational modeling for curriculum trees and prerequisite tracking
• Row-Level Security for user data isolation without custom auth infrastructure
• Mobile-first responsive design optimized for repeat engagement

Challenges & Tradeoffs

The main challenge was ensuring LLM-generated curriculum felt intentional rather than random. I solved this by designing a tree-based curriculum schema where each node has prerequisite relationships. When the correction engine detects a gap, it walks up the tree to find the true root cause, then generates targeted content from that point. The tradeoff: more complex state management, but significantly more effective curriculum delivery.

Results

What This Demonstrates

• Full-stack ownership from system design to production deployment
• RAG pipeline implementation with vector embeddings for semantic search
• AI system design beyond API usage—structured outputs, validation, and failure handling
• Performance optimization with measurable outcomes (30% API cost reduction)
• Product thinking: solving real user problems, not just building features