A complete hands-on journey from running your first local AI model to shipping a production AI SaaS product — built entirely on a Mac Mini M4.
44 weeks | 8 phases | 45+ projects | 100% local-first
Origin: This roadmap started with a viral LinkedIn post — "Build these 10 RAG projects if you want to be taken seriously as an AI engineer." Those 10 projects became Phase 2, and the journey grew from there into a full 8-phase path: foundations → RAG → agents → frameworks → multi-agent → production → advanced patterns → shipping.
| Layer | Technology |
|---|---|
| Local AI | Ollama + Gemma3:4b / 27b |
| Orchestration | LangGraph, CrewAI, LangChain |
| Vector DB | ChromaDB → Qdrant → Pinecone |
| APIs | FastAPI + Uvicorn |
| UI | Streamlit |
| Multi-Agent | LangGraph StateGraph, CrewAI Crews |
| Containers | Docker + Docker Compose |
| Cloud | AWS ECS Fargate + ECR + ALB |
| IaC | Terraform |
| Monitoring | Prometheus + Grafana |
| Fine-Tuning | LoRA + Unsloth + GGUF |
| Integrations | Slack, GitHub, Stripe, Email |
Goal: Get local AI running on Mac Mini M4
- Installed Ollama, Python 3.11, VS Code
- OpenAI-compatible local API client
- Verified full stack with gemma3:4b
📁 Phase1_Foundations/
Goal: Build the 10 hands-on Retrieval-Augmented Generation projects from the original LinkedIn post
| # | Project |
|---|---|
| 01 | First RAG pipeline (build from scratch) |
| 02 | IBM-style RAG (production patterns) |
| 03 | GraphRAG (knowledge graph) |
| 04 | Multi-document RAG (vector database) |
| 05 | Agentic RAG (autonomous agents) |
| 06 | LangChain RAG (production ready) |
| 07 | Document analysis (LLM + PDF) |
| 08 | Multimodal RAG (text + images) |
| 09 | AI research agent (automated analysis) |
| 10 | Real-time assistant (live RAG pipeline) |
Stack: nomic-embed-text, ChromaDB, numpy cosine similarity
📁 Phase2_RAG_Systems/
Goal: Master the complete agentic AI tech stack
| # | Project |
|---|---|
| 01 | Tool-calling agent (ReAct pattern) |
| 02 | Memory agent (short + long term) |
| 03 | Web scraping agent |
| 04 | Multi-tool agent |
| 05 | RAG evaluation (LLM-as-judge) |
| 06 | Agent API server |
Stack: FastAPI, mem0, BeautifulSoup, Ragas
📁 Phase3_Agentic_Stack/
Goal: Build something similar to LangChain from scratch
| # | Project |
|---|---|
| 01 | Model manager |
| 02 | Inference server (streaming + logging) |
| 03 | OpenAI-compatible API |
| 04 | Streamlit web UI |
| 05 | Custom agent framework (mini LangChain) |
| 06 | Full platform capstone |
Stack: FastAPI, Streamlit, SQLite, Typer
📁 Phase4_Agent_Framework/
Goal: Multiple specialized agents coordinating to solve complex tasks
| # | Project |
|---|---|
| 01 | Supervisor-Worker pattern (LangGraph) |
| 02 | CrewAI Research Crew (Researcher → Analyst → Writer) |
| 03 | Agent Communication Bus (asyncio pub/sub) |
| 04 | Code Generation Pipeline (review-revise loop) |
| 05 | Multi-Agent RAG with domain routing |
| 06 | Autonomous Research Pipeline with human-in-the-loop |
Stack: LangGraph, CrewAI, asyncio, Redis (optional)
📁 Phase5_Multi_Agent_Systems/
Goal: Take everything and make it production-ready, observable, secure, and deployable
| # | Project |
|---|---|
| 01 | Dockerize Everything (API + UI + Nginx) |
| 02 | Auth & RBAC (JWT + roles + API keys) |
| 03 | AWS Deployment (ECS Fargate + Terraform) |
| 04 | Observability (Prometheus + Grafana dashboards) |
| 05 | Fine-Tuning Gemma3:4b on Apple Silicon M4 |
| 06 | DocuMind — AI document intelligence SaaS (capstone) |
Stack: Docker, Terraform, AWS ECS, Prometheus, Grafana, LoRA, Unsloth
📁 Phase6_Production_Enterprise/
Goal: Go beyond basic agents into cutting-edge production patterns
| # | Project |
|---|---|
| 01 | GraphRAG (knowledge graph + relationship traversal) |
| 02 | Real-time streaming (WebSocket token streaming) |
| 03 | Long-term memory (persistent vector memory) |
| 04 | Mixture of Agents (query routing to specialists) |
| 05 | Self-improving agent (Reflexion loop) |
| 06 | AI safety & red-teaming (guardrails + adversarial tests) |
Stack: networkx, FastAPI WebSockets, SQLite, LangGraph
📁 Phase7_Advanced_AI_Patterns/
Goal: Take local-AI into real-world integrations and ship a SaaS
| # | Project |
|---|---|
| 01 | Slack bot (Socket Mode) |
| 02 | GitHub review bot (PR webhook → AI review) |
| 03 | Email agent (classify, prioritize, draft) |
| 04 | Multi-tenant SaaS (FastAPI + JWT + quotas) |
| 05 | Billing & metering (token usage + invoices + Stripe) |
| 06 | Capstone launch (dockerized RAG SaaS) |
Stack: slack-bolt, FastAPI, SQLAlchemy, Stripe, Docker
📁 Phase8_Integrations_Shipping/
# Clone
git clone https://github.com/bipinhcs11/Agentic-AI-Learning-Roadmap.git
cd Agentic-AI-Learning-Roadmap
# Setup Python environment
python3 -m venv ai-env
source ai-env/bin/activate
pip install -r requirements.txt
# Start Ollama
ollama pull gemma3:4b
ollama serve
# Run any project
python Phase2_RAG_Systems/project_01_first_rag/rag_from_scratch.pycd Phase6_Production_Enterprise/project_06_capstone_product
docker compose up --build
python demo/seed_data.py
# Open http://localhost → login: admin / admin123All projects run locally on:
- Mac Mini M4 (16GB unified memory)
- No cloud GPU required for any phase including fine-tuning
- Ollama handles model serving natively on Apple Silicon
├── Phase1_Foundations/ # setup + first model (docs/, test_gemma3.py)
├── Phase2_RAG_Systems/ # 10 RAG projects (+ guide)
├── Phase3_Agentic_Stack/ # 6 agent projects (+ guide)
├── Phase4_Agent_Framework/ # 6 framework projects (+ guide)
├── Phase5_Multi_Agent_Systems/ # 6 multi-agent projects
├── Phase6_Production_Enterprise/ # 6 production projects
├── Phase7_Advanced_AI_Patterns/ # 6 advanced-pattern projects
├── Phase8_Integrations_Shipping/ # 6 integration / shipping projects
├── scripts/ # setup & install helper scripts
├── requirements.txt # shared dependencies
├── CLAUDE.md # project context
└── README.md