Gen AI Engineer - 3-Week Intensive Learning Plan
**A structured, hands-on path from Python developer to Gen AI Engineer in 21 days.**
[Week 1: Foundations](#week-1-foundations---deep-learning-transformers--llms) | [Week 2: Core Skills](#week-2-core-genai-skills---rag-agents--chains) | [Week 3: Production](#week-3-production-skills-advanced-projects--portfolio) | [Resources](#-quick-reference-all-free-resources) | [Projects](#-project-summary)
Prerequisites
| Requirement |
Details |
| Languages |
Python (comfortable with classes, functions, data structures) |
| Libraries |
Pandas, Matplotlib, NumPy (basic usage) |
| Time |
Full day (~10-12 hours/day) for 3 weeks |
| Approach |
Learn concepts in the morning, build projects in the afternoon |
You don’t have to do this alone. The CODERCOPS Discord has a place for this roadmap.
- Run
/checkin once a day and pick Gen AI Engineer in 3 Weeks. A bot keeps a log thread with your name on it and tracks your streak.
- Ask questions in your log thread, and run
/streaks to see how everyone else is doing.
Progress Tracker
| Week |
Day |
Topic |
Status |
| 1 |
Day 1 |
Neural Networks & Backpropagation |
✅ |
| 1 |
Day 2 |
Language Modeling & NLP Basics |
✅ |
| 1 |
Day 3 |
Transformers - The Core of All GenAI |
✅ |
| 1 |
Day 4 |
Tokenization & Hugging Face Basics |
✅ |
| 1 |
Day 5 |
Prompt Engineering & LLM APIs |
⬜ |
| 1 |
Day 6 |
Embeddings & Vector Databases |
⬜ |
| 1 |
Day 7 |
Review & Consolidation |
⬜ |
| 2 |
Day 8 |
LangChain Fundamentals |
⬜ |
| 2 |
Day 9 |
RAG - Basic Implementation |
⬜ |
| 2 |
Day 10 |
RAG - Advanced + YouTube Project |
⬜ |
| 2 |
Day 11 |
AI Agents - Tool Use & ReAct |
⬜ |
| 2 |
Day 12 |
LangGraph & Multi-Step Agents |
⬜ |
| 2 |
Day 13 |
Fine-Tuning LLMs |
⬜ |
| 2 |
Day 14 |
Review, Refine & Integrate |
⬜ |
| 3 |
Day 15 |
Multi-Agent Systems |
⬜ |
| 3 |
Day 16 |
Agentic RAG & Advanced Patterns |
⬜ |
| 3 |
Day 17 |
Deployment & APIs |
⬜ |
| 3 |
Day 18 |
Evaluation & Observability |
⬜ |
| 3 |
Day 19 |
Multimodal AI & Vision |
⬜ |
| 3 |
Day 20 |
Ethics, Safety & Responsible AI |
⬜ |
| 3 |
Day 21 |
Portfolio Day - Final Polish |
⬜ |
Understand how neural networks and LLMs actually work under the hood. Without this, everything else is just API calls with no understanding.
| Morning (4 hrs) - Learn |
|
- Andrej Karpathy - [The spelled-out intro to neural networks and backpropagation: building micrograd](https://www.youtube.com/watch?v=VMj-3S1tku0)
- Follow along and code micrograd yourself
|
| Afternoon (4 hrs) - Build |
|
- Rebuild micrograd from scratch without looking at the video
- Extend it: add `tanh`, `relu`, `sigmoid` activation functions
- Train a small classifier on a toy dataset
|
| Resources |
|
- [micrograd GitHub repo](https://github.com/karpathy/micrograd)
|
| Morning (5 hrs) - Learn |
|
- Karpathy - [The spelled-out intro to language modeling: building makemore](https://www.youtube.com/watch?v=PaCmpygFfXo)
- Karpathy - [Building makemore Part 2: MLP](https://www.youtube.com/watch?v=TCH_1BHY58I)
|
| Afternoon (4 hrs) - Build |
|
- Build a character-level name generator from scratch
- Train it on a dataset of names
- Experiment: change hidden layer sizes, learning rates, plot loss curves with matplotlib
|
| Resources |
|
- [makemore GitHub repo](https://github.com/karpathy/makemore)
|
| Morning (5 hrs) - Learn |
|
- Karpathy - [Let's build GPT: from scratch, in code, spelled out](https://www.youtube.com/watch?v=kCc8FmEb1nY)
- Jay Alammar - [The Illustrated Transformer](https://jalammar.github.io/illustrated-transformer/)
|
| Afternoon (5 hrs) - Build |
|
- Code a mini GPT from scratch following Karpathy's video
- Train it on a text file of your choice (Shakespeare, song lyrics, etc.)
- Generate text and observe how quality changes with training
|
| Resources |
|
- [nanoGPT GitHub repo](https://github.com/karpathy/nanoGPT)
|
| Morning (4 hrs) - Learn |
|
- Karpathy - [Let's build the GPT Tokenizer](https://www.youtube.com/watch?v=zduSFxRajkE)
- [Hugging Face LLM Course - Chapter 1](https://huggingface.co/learn/llm-course/chapter1/1)
|
| Afternoon (5 hrs) - Build |
|
- Build a BPE tokenizer from scratch
- Use Hugging Face `transformers` library to load pre-trained models (GPT-2)
- Generate text, experiment with temperature and top-k/top-p sampling
- Compare outputs: GPT-2, DistilGPT-2, GPT-2-medium
|
pip install transformers datasets accelerate tiktoken
| Morning (4 hrs) - Learn |
|
- DeepLearning.AI - [ChatGPT Prompt Engineering for Developers](https://www.deeplearning.ai/short-courses/chatgpt-prompt-engineering-for-developers/) (free, ~1.5 hrs)
- DeepLearning.AI - [Building Systems with the ChatGPT API](https://www.deeplearning.ai/short-courses/building-systems-with-chatgpt/) (free, ~1.5 hrs)
|
| Afternoon (6 hrs) - Build |
|
**Project 1: AI-Powered Resume Analyzer** - A Python CLI tool that takes a resume as input, extracts skills/experience/education, scores against a job description, suggests improvements, and generates a cover letter draft.
|
| Morning (4 hrs) - Learn |
|
- DeepLearning.AI - [Building Applications with Vector Databases](https://www.deeplearning.ai/short-courses/building-applications-vector-databases/) (free)
- Vicki Boykis - [What are Embeddings?](https://vickiboykis.com/what_are_embeddings/)
|
| Afternoon (5 hrs) - Build |
|
**Project 2: Semantic Search Engine for Your Notes** - Collect 50+ text files, generate embeddings with `sentence-transformers/all-MiniLM-L6-v2` (runs locally), store in ChromaDB, build a CLI for semantic search with cross-encoder re-ranking.
|
pip install chromadb sentence-transformers
| Morning (4 hrs) |
|
- Review all code from the week
- Revisit any weak concepts
- Read: [Hugging Face LLM Course Chapters 2-3](https://huggingface.co/learn/llm-course/chapter2/1)
|
| Afternoon (5 hrs) |
|
- Clean up Project 1 and Project 2
- Add proper error handling, README files, push to GitHub
- Write a short blog post / notes about what you learned
|
Week 2: Core GenAI Skills - RAG, Agents & Chains
Learn the bread-and-butter of a Gen AI Engineer: building applications that combine LLMs with external data and tools.
| Morning (5 hrs) - Learn |
|
- [LangChain Official Tutorials](https://python.langchain.com/docs/tutorials/):
- Build a Simple LLM Application
- Build a Chatbot
- Build a RAG app
|
| Afternoon (5 hrs) - Build |
|
- Set up a LangChain project from scratch
- Build chains: simple -> sequential -> chain with memory
- Experiment with different LLMs (OpenAI, Groq/Llama, Gemini)
- Add conversation memory (BufferMemory, SummaryMemory)
|
pip install langchain langchain-openai langchain-community
| Day 9 - Learn & Build Basic RAG (10 hrs) |
|
- DeepLearning.AI - [LangChain: Chat with Your Data](https://www.deeplearning.ai/short-courses/langchain-chat-with-your-data/) (free)
- Hugging Face Blog - [Code a simple RAG from scratch](https://huggingface.co/blog/ngxson/make-your-own-rag)
- [LangChain RAG Tutorial](https://python.langchain.com/docs/tutorials/rag/)
|
| Day 10 - Build Advanced RAG (10 hrs) |
|
**[Project 3: Chat With Any YouTube Channel](week2/day10_rag_advanced_youtube_project/)** - Download transcripts of last 20 videos, chunk intelligently (256, 512, 1024), embed and store in ChromaDB, build Q&A with source citations (video + timestamp), Streamlit UI.
|
Key RAG concepts to implement:
| Concept |
Description |
| Document Loading |
Text splitting strategies for different content types |
| Embedding Models |
OpenAI embeddings vs local sentence-transformers |
| Vector Store |
ChromaDB for storage and retrieval |
| Retrieval Strategies |
Similarity search, MMR (Maximum Marginal Relevance) |
| Prompt Templates |
QA with context injection |
pip install youtube-transcript-api streamlit chromadb langchain
| Morning (5 hrs) - Learn |
|
- DeepLearning.AI - [Functions, Tools and Agents with LangChain](https://www.deeplearning.ai/short-courses/functions-tools-agents-langchain/) (free)
- Microsoft - [AI Agents for Beginners, Lessons 1-4](https://github.com/microsoft/ai-agents-for-beginners)
- [ReAct Pattern](https://www.promptingguide.ai/techniques/react)
|
| Afternoon (5 hrs) - Build |
|
- Build an agent with LangChain that can search the web, do math, read/write files, and execute Python code
- Understand the **Thought -> Action -> Observation** loop
|
pip install langchain-community duckduckgo-search
| Morning (4 hrs) - Learn |
|
- [LangChain Academy - Intro to LangGraph](https://academy.langchain.com/courses/intro-to-langgraph) (free)
- [LangGraph Tutorials](https://langchain-ai.github.io/langgraph/tutorials/)
|
| Afternoon (6 hrs) - Build |
|
**Project 4: AI Research Agent** - An agent using LangGraph that generates search queries, searches the web, summarizes results, synthesizes a research report, and identifies follow-up questions. Graph-based workflow with conditional edges.
|
pip install langgraph langchain-openai tavily-python
| Morning (5 hrs) - Learn |
|
- DeepLearning.AI - [Finetuning Large Language Models](https://www.deeplearning.ai/short-courses/finetuning-large-language-models/) (free)
- [Hugging Face LLM Course - Chapter 11](https://huggingface.co/learn/llm-course/chapter11/1)
- [LoRA and QLoRA Guide](https://huggingface.co/docs/peft/conceptual_guides/lora)
|
| Afternoon (5 hrs) - Build |
|
**Project 5: Fine-Tune a Model for Your Use Case** - Pick a task (sentiment analysis, code explanation, domain Q&A), apply LoRA adapters to a small model (`phi-2` or `TinyLlama`), evaluate before vs after. Use Google Colab free tier if no GPU.
|
pip install peft transformers datasets accelerate bitsandbytes
| Morning (4 hrs) |
|
- Review Week 2 projects
- Try different chunking strategies, compare retrieval methods
- Read: [NirDiamant's GenAI Agents](https://github.com/NirDiamant/GenAI_Agents) for more patterns
|
| Afternoon (5 hrs) |
|
- Improve Project 3 (RAG): hybrid search, query rewriting, simple evaluation with 10 Q&A pairs
- Push all projects to GitHub with good READMEs
|
Week 3: Production Skills, Advanced Projects & Portfolio
Build production-grade projects and learn deployment. This is what separates a learner from a hirable engineer.
| Morning (4 hrs) - Learn |
|
- DeepLearning.AI - [Multi AI Agent Systems with crewAI](https://www.deeplearning.ai/short-courses/multi-ai-agent-systems-with-crewai/) (free)
- Browse: [500 AI Agent Projects](https://github.com/ashishpatel26/500-AI-Agents-Projects) for ideas
|
| Afternoon (6 hrs) - Build |
|
**Project 6: AI Content Creation Pipeline** - Multi-agent system using CrewAI: Researcher (gathers facts), Writer (drafts blog post), Editor (reviews & improves), SEO Optimizer (keywords, meta descriptions). Output: publish-ready blog post.
|
pip install crewai crewai-tools
| Morning (4 hrs) - Learn |
|
- DeepLearning.AI - [Building Agentic RAG with LlamaIndex](https://www.deeplearning.ai/short-courses/building-agentic-rag-with-llamaindex/) (free)
- [DataCamp Agentic RAG Tutorial](https://www.datacamp.com/tutorial/agentic-rag-tutorial)
|
| Afternoon (6 hrs) - Build |
|
**Project 7: AI Coding Assistant** - Agentic RAG that indexes a GitHub repo's codebase, answers questions about it, navigates between files following imports, and uses an agent loop to reformulate queries when first retrieval isn't sufficient.
|
pip install llama-index llama-index-vector-stores-chroma
| Morning (3 hrs) - Learn |
|
- [FastAPI Tutorial](https://fastapi.tiangolo.com/tutorial/)
- [Streamlit Docs](https://docs.streamlit.io/get-started)
- [Docker Get Started](https://docs.docker.com/get-started/)
|
| Afternoon (7 hrs) - Build |
|
**Project 8: Deploy Your Best RAG App as a Full Product** - FastAPI backend (`POST /upload`, `POST /ask`, `GET /history`), Streamlit frontend, SQLite for conversation memory, streaming responses, Docker, deploy on Streamlit Cloud / HF Spaces / Railway.
|
| Morning (4 hrs) - Learn |
|
- DeepLearning.AI - [Building and Evaluating Advanced RAG](https://www.deeplearning.ai/short-courses/building-evaluating-advanced-rag/) (free)
- [RAGAS Documentation](https://docs.ragas.io/en/latest/)
- [DeepEval Documentation](https://docs.confident-ai.com/)
|
| Afternoon (5 hrs) - Build |
|
- Create a test dataset of 20+ question-answer pairs
- Measure: faithfulness, answer relevancy, context precision, context recall
- Use RAGAS framework, generate evaluation reports
- Iterate: change chunk size, embedding model, prompt template and re-evaluate
|
pip install ragas deepeval
| Morning (4 hrs) - Learn |
|
- DeepLearning.AI - [Prompt Engineering for Vision Models](https://www.deeplearning.ai/short-courses/prompt-engineering-for-vision-models/) (free)
- Explore: GPT-4o (vision API), LLaVA (open source), Google Gemini (free multimodal API)
|
| Afternoon (6 hrs) - Build |
|
**Project 9: AI Invoice/Receipt Processor** - Upload photos of invoices/receipts, extract vendor name, date, total, line items, tax using a vision model, store in SQLite, generate monthly expense reports with charts, Streamlit UI.
|
| Morning (3 hrs) - Learn |
|
- [Generative AI for Everyone by Andrew Ng](https://www.coursera.org/learn/generative-ai-for-everyone) (free on Coursera, focus Weeks 2-3)
- Topics: hallucination mitigation, prompt injection attacks & defenses, bias in LLMs, EU AI Act basics
|
| Afternoon (5 hrs) - Build |
|
- Add safety features: input validation, prompt injection detection, output guardrails, content filters
- Red-team your own chatbot: try to make it produce incorrect outputs
|
| Morning (5 hrs) |
|
- Pick your **3 best projects** and polish them
- Clean code, proper docstrings, comprehensive READMEs with screenshots/GIFs
- Architecture diagrams, tech stack documentation
|
| Afternoon (5 hrs) |
|
- Create a portfolio page (GitHub profile README or simple site)
- Write a LinkedIn post about your 3-week journey
- Plan continued learning: communities, people to follow, open source contributions
|
📚 Quick Reference: All Free Resources
Video Courses
Documentation
GitHub Repos
Free API Keys
Free Compute (for fine-tuning)
🏗 Project Summary
| # |
Project |
Day |
Key Skills |
Difficulty |
| 1 |
AI Resume Analyzer |
5 |
LLM APIs, prompt engineering, structured output |
⭐⭐ |
| 2 |
Semantic Search Engine |
6 |
Embeddings, vector DB, sentence-transformers |
⭐⭐ |
| 3 |
Chat With YouTube Channel |
9-10 |
RAG, chunking, ChromaDB, Streamlit |
⭐⭐⭐ |
| 4 |
AI Research Agent |
12 |
LangGraph, web search, multi-step reasoning |
⭐⭐⭐ |
| 5 |
Fine-Tuned Domain Model |
13 |
Hugging Face, LoRA, PEFT, evaluation |
⭐⭐⭐ |
| 6 |
AI Content Pipeline |
15 |
Multi-agent, CrewAI, tool use |
⭐⭐⭐⭐ |
| 7 |
AI Coding Assistant |
16 |
Agentic RAG, code understanding |
⭐⭐⭐⭐ |
| 8 |
Deployed RAG Product |
17 |
FastAPI, Docker, Streamlit, deployment |
⭐⭐⭐⭐ |
| 9 |
Invoice/Receipt Processor |
19 |
Multimodal AI, vision, structured extraction |
⭐⭐⭐⭐ |
Setup
Environment
# Install Miniconda (if not installed)
# https://docs.conda.io/en/latest/miniconda.html
# Create and activate environment
conda create -n genai python=3.11 -y
conda activate genai
# Install core packages
conda install pytorch matplotlib numpy scikit-learn -y
pip install transformers datasets langchain chromadb sentence-transformers
Repository Structure
gen_ai_engineer/
├── README.md # This file
├── week1/
│ ├── day01_neural_networks_backpropagation/
│ │ ├── README.md # Book: Neural Networks & Backprop
│ │ ├── micrograd.py # Autograd engine + MLP
│ │ ├── exercises.py # Hands-on exercises
│ │ ├── visualizations.py # Generate 8 concept diagrams
│ │ └── viz_*.png # Pre-generated diagrams
│ ├── day02_language_modeling_nlp/
│ │ ├── README.md # Book: Language Modeling & NLP
│ │ ├── bigram.py # Counting-based bigram model
│ │ ├── bigram_neural.py # Neural network bigram
│ │ ├── mlp_lm.py # MLP language model
│ │ └── names.txt # Dataset (32K names)
│ ├── day03_transformers/
│ │ ├── README.md # Book: Transformers
│ │ ├── mini_gpt.py # Decoder-only Transformer LM
│ │ ├── visualizations.py # Generate 8 transformer diagrams
│ │ └── viz_*.png # Pre-generated diagrams
│ ├── day04_tokenization_huggingface/
│ │ ├── README.md # Book: Tokenization & HF
│ │ ├── bpe_tokenizer.py # BPE tokenizer from scratch
│ │ ├── huggingface_generate.py # GPT-2 loading & generation
│ │ ├── visualizations.py # Generate 8 concept diagrams
│ │ └── viz_*.png # Pre-generated diagrams
│ └── day05-07 .../
├── week2/
│ └── day08-14 .../
└── week3/
└── day15-21 .../
> **The goal is not to finish everything perfectly. It's to build muscle memory for the GenAI stack.**
> A working ugly project teaches more than a polished tutorial you followed.
> When stuck, read the error, check the docs, ask an LLM - that IS the job of a GenAI engineer.