Scaffi AI
Smart Study Buddy for Programming Students
Tech Stack
Links
The Problem
Learning to code is hard. And the way students learn today is broken.
ChatGPT Dependency
78% of students use AI to copy-paste answers without understanding, failing exams and interviews.
Analysis Paralysis
Students freeze when facing large, complex projects because they lack the roadmap to break them down.
The Educational Gap
Students need something between "complete answer" (ChatGPT) and "figure it out yourself" (textbooks).
Solution
Scaffi-AI is like a GPS for coding homework. Instead of just showing you the destination, it gives you turn-by-turn directions, checks your progress, and helps when you're stuck.
1. Upload Assignment
Students upload strict PDF requirements. Supports 10+ languages (Java, C++, Python, etc.).
2. AI Decomposition
AI breaks the problem into 8-15 manageable tasks with estimated times (e.g., "Create Producer Class - 10 min").
3. Progressive Hints
Stuck? Get hints that nudge you: Conceptual → Specific → Pseudocode pattern. Never the full answer.
4. Smart Testing
Generates tests based on student code logic, ensuring creative solutions get full credit if valid.
How I Built It
Multi-Agent AI Architecture
Parser Agent (Claude Haiku)
Reads PDFs, detects multi-file projects, and extracts class structures cost-effectively.
Code Gen Agent (Claude Sonnet)
Creates scaffolded starter code. Preserves professor's templates while adding dynamic TODOs.
Helper & Concept Agents
Provides progressive hints and context-aware analogies (e.g., explaining C++ pointers using Python references).
Key Technical Achievements
- Template Preservation Engine Regex-based extraction ensures student code aligns exactly with professor-provided starter files.
- Pattern-Based Test Matching Handles non-deterministic outputs (threading/randomness) using "CONTAINS" and "COUNT" fuzzy matching.
- Session Recovery LocalStorage isolation with 500ms debounce auto-save. Survives browser crashes with 0 data loss.