MOSAiC: AI Personalized Learning Platform

The Problem

Most online learning is one-size-fits-all: the same modules in the same order whether you know the material or not. MOSAiC generates personalized paths instead.

The content rarely adapts to how you actually learn. MOSAiC was built to fix that, an AI-powered platform that tailors what you see rather than serving everyone the same fixed course. I built the full stack as CTO.

What It Does

MOSAiC pairs an adaptive learning engine with an AI guide called Tessera that walks learners through material, answers questions, and adjusts to where they are.

Goal tracking keeps a learner oriented, multimodal content mixes formats rather than dumping walls of text, and a quiz system checks understanding and feeds back into the adaptive engine. On the platform it had generated 12,000+ personalized modules and pulled in 2,000+ waitlist signups.

The Architecture

The platform runs on Next.js with a multi-agent AI architecture and a knowledge graph that structures how concepts relate so material sequences sensibly.

The interesting decision was making module generation adaptive rather than pre-authored. Generating content on demand makes the catalog effectively unbounded, but it puts real weight on quality control, since a bad generated module reaches a learner directly. Most of my engineering time went into keeping generated content accurate and coherent, not into the UI.

Building It With AI

I leaned hard on AI coding tools like Cursor and Claude Code to move fast as a small team, which is a running theme across everything I build.

There is a certain recursion in using AI agents to build an AI-powered learning product, and the same lesson kept surfacing: the models are great at generating, and the hard part is the judgment about what is actually good. You can read more on that mindset in my Claude Code guide.

What I'd Do Differently

The biggest ongoing tension in a generative learning product is trust, so I would invest earlier in evaluation tooling.

Automatically scoring generated modules for accuracy and difficulty before they ever reach a learner beats catching problems reactively. If the founder-tooling and edtech angle interests you, I also built SysDesignLab for hands-on system design practice and AIdeaHub for taking an idea to a structured launch plan.