LearnEngine

Type literally any topic and get a structured lesson plan in seconds. The AI breaks it down into 4-6 sections with time estimates. Getting different AI models to format their output consistently was... an adventure.

• •OpenRouter integration, had to debug each model separately because they all have quirks
• •Mimo Flash, GPT-120B, GLM 4.5 Air support (each returns slightly different JSON, fun times)
• •Prompt injection sanitization after someone tried some creative inputs
• •JSON parsing with fallbacks because AI outputs aren't always clean

Each lesson gets detailed, AI-written content in Markdown with math equation support. Getting the difficulty slider to actually change the language complexity took way more prompt iterations than I want to admit.

• •400-600 words per section (had to tweak prompts to stop AI from writing novels)
• •Difficulty scale 1-10 that actually works (mostly)
• •Markdown rendering with syntax highlighting for code examples
• •KaTeX for math equations, surprisingly smooth to integrate

Questions get easier if you're struggling (<50% accuracy) or harder if you're crushing it (>80%). The algorithm tracks your last 10 answers. First version was way too aggressive with difficulty swings, had to tune it.

• •Multiple choice, true/false, and short answer question types
• •AI grading for open-ended responses (this took forever to get right)
• •Performance context tracking that actually influences new questions
• •Difficulty adjustment in -2 to +2 range, aggressive enough to feel adaptive, not so much it's jarring

Built-in AI tutor that knows what lesson you're on. Ask questions anytime and get contextual help. The first version lost chat history on page refresh, that was a fun bug to explain to testers.

• •Persistent chat history that actually persists (took a few attempts)
• •Model picker so you can choose your AI
• •Lesson context injection so the AI knows what you're learning
• •Message streaming for that ChatGPT-like typing feel

Tracks accuracy, time spent, and improvement over time. Everything persists across sessions. The database schema for this went through like 3 revisions before it made sense.

• •Per-question answer tracking (yes, I store every answer)
• •Time spent per section and per question
• •Accuracy calculations that update in real-time
• •Visual progress bars that actually feel motivating

Multi-layer security because I'm paranoid. Rate limiting, input sanitization, RLS policies, the works. Found an open redirect vulnerability during my own testing which was embarrassing but better than finding it in prod.

• •Supabase Auth with proper session handling
• •Rate limiting at 10 AI calls/minute (my wallet said so)
• •Prompt injection pattern filtering, people will try to break things
• •Open redirect prevention after I found a vulnerability in my own code
• •RLS policies that actually work (after the semicolon incident)

PostgreSQL with 7 interconnected tables. Users, lessons, sections, questions, progress, chat history, the whole thing. Took a while to get the relationships right but now it's actually pretty clean.

• •Profiles, lesson_plans, lesson_sections tables (basic but solid)
• •Questions with difficulty levels and type metadata
• •User progress tracking per section and per question
• •Chat conversation and message storage with proper foreign keys
• •RLS policies on everything, nobody sees anyone else's data