learnings 7 min read EN

AI PE Learning Agent Project Setup

Building AI agents to learn AI engineering - the tool creation process itself becomes the learning journey. Instead of following tutorials, you solve...

#meta-learning #multi-agent #langgraph #project-setup #portfolio
Table of Contents

Complete project initialization for AI Product Engineer learning journey with meta-learning approach

Date: 2026-01-18 Category: AI Engineering, Project Management Tags: #meta-learning #multi-agent #langgraph #project-setup #portfolio

Key Concepts

Meta-Learning Approach

Building AI agents to learn AI engineering - the tool creation process itself becomes the learning journey. Instead of following tutorials, you solve a real problem (managing your own learning) while acquiring AI engineering skills.

Multi-Agent System Architecture

Four specialized agents working together:

  • Curriculum Architect: Curates weekly learning goals and resources
  • Assignment Generator: Creates practical coding assignments
  • Code Critic: Reviews code for AI Engineering Best Practices
  • Progress Tracker: Manages learning history in Vector DB

Separate Repository Strategy

ai-pe-learning-agent as standalone project, not a plugin in claude-ai-engineering:

  • Different scope and purpose
  • Independent evolution and versioning
  • Clear portfolio value
  • Dedicated learning artifacts in learning/ directory

Learning Artifacts Structure

learning/
├── week-00-python-basics/
│   ├── notes.md              # Learning notes
│   ├── exercises/            # Practice code
│   ├── assignments/          # Agent-generated tasks
│   └── reviews/              # Code review results

This structure proves:

  • Transparent learning process
  • Dogfooding (using your own tools)
  • Growth over time (failure → improvement)
  • Real usage examples

New Learnings

Before Understanding

  • Initially considered including AI PE project in plugins/ directory
  • Thought documents should be in Korean for personal use
  • Unclear about repository structure and organization
  • Uncertain about what to include in learning artifacts

After Understanding

Repository Organization:

  • Separate repositories have clear boundaries and purposes
  • claude-ai-engineering: Plugin marketplace and AI engineering toolkit
  • ai-pe-learning-agent: Dedicated learning management system

Documentation Strategy:

  • English for broader reach and portfolio value
  • Personal information (PROFILE.md) excluded from git commits
  • Origin documents preserved (PROJECT-ORIGIN.md from idea.md)
  • Comprehensive roadmap with 5 phases + Week 0 Python basics

Learning Process:

  • Week 0 added for TypeScript → Python transition (1 week)
  • Each week includes notes, exercises, assignments, and reviews
  • Learning artifacts are core portfolio differentiator
  • Failed attempts documented alongside successes

Tech Stack Decision:

  • Python-first approach (AI ecosystem standard)
  • TypeScript optional for later productization
  • LangGraph for agentic workflows (not entire LangChain)
  • Focus on depth over breadth

Practical Examples

Project Structure Created

ai-pe-learning-agent/
├── README.md                    # English
├── docs/
│   ├── ROADMAP.md              # Comprehensive 5-phase plan
│   ├── GETTING-STARTED.md      # Setup guide
│   ├── PROJECT-ORIGIN.md       # Genesis of idea
│   └── PROFILE.md              # Personal info (gitignored)
├── src/
│   ├── __init__.py
│   ├── hello_claude.py         # First example with TS comparisons
│   └── agents/
│       └── __init__.py
├── learning/
│   └── week-00-python-basics/
│       └── notes.md            # TypeScript → Python guide
├── requirements.txt
├── .env.example
└── .gitignore

Git Ignore Pattern for Personal Data

# Personal information (excluded from commits)
docs/PROFILE.md

This allows agents to use your background for personalization while keeping it private.

TypeScript to Python Comparison Pattern

# TypeScript
const numbers = [1, 2, 3, 4, 5];
const doubled = numbers.map(n => n * 2);

# Python
numbers = [1, 2, 3, 4, 5]
doubled = [n * 2 for n in numbers]  # List comprehension!

Week 0 notes.md provides extensive comparisons for TypeScript developers.

Repository Naming

Repository: ai-pe-learning-agent
Description: Multi-agent system for AI Product Engineer learning journey.
             Meta-learning project: building AI agents to become an AI
             engineer. LangGraph, Claude API, Vector DB, Python.

Clear, searchable, and conveys purpose immediately.

Common Misconceptions

”Learning artifacts should be private”

Reality: Public learning process is a portfolio strength. Shows:

  • Transparent growth journey
  • Willingness to share mistakes and improvements
  • Real-world problem-solving
  • Consistent learning habit

Exception: Personal career details (PROFILE.md) excluded via .gitignore

”Documentation should be in native language for personal projects”

Reality: English documentation:

  • Increases portfolio reach
  • Demonstrates communication skills
  • Makes project accessible to global audience
  • Better for job applications and networking

”Need to learn entire LangChain ecosystem”

Reality: Focus on LangGraph specifically:

  • State-based workflows
  • Circular structures (retry logic)
  • Visualization and debugging
  • Avoid scope creep by learning selectively

”Should include everything in one repository”

Reality: Separate repositories when:

  • Different scope and purpose
  • Independent evolution needed
  • Clear boundary between projects
  • Better organization and clarity

References

Files Created/Modified

New ai-pe-learning-agent project:

  • /Users/jaykim/Documents/Projects/ai-pe-learning-agent/README.md
  • /Users/jaykim/Documents/Projects/ai-pe-learning-agent/docs/ROADMAP.md (38,237 bytes)
  • /Users/jaykim/Documents/Projects/ai-pe-learning-agent/docs/GETTING-STARTED.md
  • /Users/jaykim/Documents/Projects/ai-pe-learning-agent/docs/PROJECT-ORIGIN.md
  • /Users/jaykim/Documents/Projects/ai-pe-learning-agent/docs/PROFILE.md (gitignored)
  • /Users/jaykim/Documents/Projects/ai-pe-learning-agent/learning/week-00-python-basics/notes.md
  • /Users/jaykim/Documents/Projects/ai-pe-learning-agent/src/hello_claude.py
  • /Users/jaykim/Documents/Projects/ai-pe-learning-agent/.gitignore
  • /Users/jaykim/Documents/Projects/ai-pe-learning-agent/requirements.txt

Removed from claude-ai-engineering:

  • docs/ folder (AI-PE-ROADMAP.md, AI-PE-ROADMAP2.md, ROADMAP-COMPARISON.md)

Source documents:

  • idea.md → PROJECT-ORIGIN.md (translated)
  • career.md → PROFILE.md (translated, gitignored)
  • test.md → insights integrated into ROADMAP.md

Key Documentation

ROADMAP.md sections:

  • Phase 0: Project Preparation (1 week)
  • Phase 1: MVP - Prompt Reviewer (2 weeks) - Few-shot, CoT, ReAct
  • Phase 2: Code Reviewer (2 weeks) - Validation, Security
  • Phase 3: Vector DB Integration (2 weeks) - RAG, Hybrid Search
  • Phase 4: LangGraph + Evaluation (2 weeks) - LLM-as-a-Judge
  • Phase 5: Full Integration + Optimization (2 weeks) - Semantic Caching

Week 0 Content:

  • Day 1-2: Python Fundamentals (TypeScript comparisons)
  • Day 3-4: AI Ecosystem (pip, venv, Claude API)
  • Day 5: Pydantic (Data Validation)
  • Day 6-7: Mini Chatbot Project

Technologies Mentioned

  • Languages: Python 3.11+, TypeScript (future)
  • AI: Claude API (Anthropic SDK), GPT-4o (for LLM-as-a-Judge)
  • Frameworks: LangGraph, LangChain (selective)
  • Storage: Vector DB (Supabase/Chroma), JSON files
  • Tools: Pydantic, Poetry/uv, pytest, ruff, mypy
  • Optional: Streamlit, Next.js, Docker

Decision Points

1. Repository Structure

Decision: Separate repository (ai-pe-learning-agent) Rationale:

  • Different scope from claude-ai-engineering toolkit
  • Clear portfolio boundary
  • Independent versioning and evolution
  • Dedicated learning artifacts directory

2. Language Strategy

Decision: Python-first, TypeScript optional later Rationale:

  • AI ecosystem centers on Python
  • LangChain/LangGraph Python-native
  • Can add TypeScript for productization in Phase 5
  • Week 0 bridges TypeScript → Python transition

3. Roadmap Choice

Decision: Python “정석” (standard) approach over TypeScript version Rationale:

  • Industry standard (50/50 score vs 30/50)
  • Deeper AI engineering fundamentals
  • Better for long-term career growth
  • TypeScript skills still valuable for future integration

4. Documentation Language

Decision: All English Rationale:

  • Portfolio and job application value
  • Global reach
  • Professional communication demonstration
  • Personal info (PROFILE.md) excluded via gitignore

5. Learning Artifacts Visibility

Decision: Public by default, personal info excluded Rationale:

  • Transparent learning process shows growth
  • Dogfooding evidence (using own tools)
  • Interview preparation material
  • Blog post source material

Portfolio Value

Quantitative Metrics Planned

From ROADMAP.md:

  • Learning Records: 12 weeks, 50+ assignments, 100+ code reviews
  • Performance Improvements:
    • Prompt quality: 40% → 85% (LLM-as-a-Judge measured)
    • Response consistency: 60% → 99% (Guardrails)
    • Hallucination: 25% → 0% (Validation logic)
  • Cost Reduction: API cost 70% via Semantic Caching
  • Code: Python 5,000+ lines, 80%+ test coverage
  • Blog Posts: 12+ technical posts over 6 months

Interview Preparation

  • “Why did you build this?” → PROJECT-ORIGIN.md
  • “Biggest technical challenge?” → learning/week-N/notes.md reflections
  • “Failure and recovery?” → reviews/ directory (failed → passed examples)
  • “What did you learn?” → 12 weeks of documented learning

Blog Post Topics (12+ planned)

  1. Project planning and meta-learning approach
  2. System Prompt design process
  3. Guardrails for 100% response control
  4. Code review agent with Best Practices
  5. Reducing hallucination to 0%
  6. Vector DB for learning history
  7. LLM-as-a-Judge evaluation
  8. LangGraph workflows
  9. Semantic Caching 70% cost reduction
  10. Multi-agent collaboration
  11. 12-week learning data insights
  12. 3-month retrospective

Next Steps

Immediate Actions

  • Create GitHub repository ai-pe-learning-agent
  • Push initial commit to GitHub
  • Setup Python virtual environment
  • Test Claude API with hello_claude.py
  • Complete Week 0 Python basics (7 days)

Week 0 Goals (2026-01-18 ~ 2026-01-25)

  • Understand basic Python syntax
  • Virtual environment and package management
  • Successfully call Claude API
  • Data validation with Pydantic
  • Ready to start Phase 1

Phase 1 Preparation

  • Review ROADMAP.md Phase 1 details
  • Study Few-shot, Chain-of-Thought, ReAct patterns
  • Understand Structured Output with Pydantic
  • Research Guardrails implementation
  • Write Phase 1 detailed design document

Documentation Tasks

  • Create ARCHITECTURE.md (system design details)
  • Create DEVELOPMENT.md (development log)
  • Consider blog platform (Medium, personal blog, velog)
  • Decide repository publicity strategy (public from start recommended)

Learning Resources to Review

Session Summary

What Was Accomplished:

  1. ✅ Created complete ai-pe-learning-agent project structure
  2. ✅ Translated all documentation to English (README, ROADMAP, GETTING-STARTED, PROJECT-ORIGIN)
  3. ✅ Created PROFILE.md for agent personalization (gitignored)
  4. ✅ Setup Week 0 Python basics guide with TypeScript comparisons
  5. ✅ Configured .gitignore to exclude personal information
  6. ✅ Removed docs/ folder from claude-ai-engineering
  7. ✅ Made initial git commit (11 files, 1,952 insertions)

Files Created: 11 files across project structure Total Lines: 1,952 lines of documentation and code Time Investment: Initial setup complete, ready for Week 0

Key Insight: The meta-learning approach—building AI agents to learn AI engineering—creates a self-reinforcing learning cycle. The project serves simultaneously as:

  • Learning tool (manages your own learning)
  • Learning process (builds AI engineering skills)
  • Learning artifact (portfolio evidence)

This triple-value proposition makes it a uniquely powerful portfolio project.

Created: 2026-01-18 Session Duration: Full project setup and planning Next Session: GitHub repository creation and Week 0 Python basics