nursnaaz/zero-to-genai-engineer

---

New sessions drop every Saturday / Sunday. Star the repo to get notified. Questions → WhatsApp cohort group.

---

Table of contents

1. Start here
2. Prerequisites
3. What’s in this repo
4. The MISSING chain
5. Your instructor
6. How this repo is laid out
7. Sessions shipped so far
8. Full syllabus
9. Projects
10. Classroom presentations
11. Tech stack
12. Module map
13. Contributing
14. License

---

🚀 Start here

Beginner notebooks (no install):

Then keep the Cheat Sheet open during S00.

---

⚙️ Prerequisites

git clone https://github.com/nursnaaz/zero-to-genai-engineer.git
cd zero-to-genai-engineer

Then open that weekend’s folder README. That’s the start page. There is no root requirements.txt; each session and app has its own (10_RAG/notebooks/requirements.txt, 11_LangGraph/notebooks/requirements.txt, and so on).

Copy .env.example to .env. Don’t commit the real one.

---

📦 What’s in this repo

Every weekend is notebooks you actually run. Most also leave you with an app, a paper, or a dataset you can talk about in an interview.

Quick naming trap: folder 10_RAG/ is session S10, which covers modules M07 + M08 + M06 (memory went into the RAG chatbot, not its own weekend). 11_LangGraph/ is session S11 = module M10. 12_deepagents/ is S12. 13_Project_Implementation/ is S13 (Dining Bot). 14_Bedrock_AgentCore/ is S14 — six sub-labs numbered 01. → 06.. 15_LLM_Fine_Tuning/ is S15 (LoRA / QLoRA). 16_Full_Course_Recap/ recaps S00–S15 and does not replace S08. 17_Decision_Models/ is S17. 18_Final_Project/ is S18. S14 goes beyond the original 23-module syllabus (see Module map).

---

🔗 The MISSING chain

Every session exists because the last one left a gap. That chain is how the course is sequenced.

S00  Search (TF-IDF)
  └─ MISSING: meaning  ─►  S01  Embeddings (BoW → FastText)
       └─ MISSING: context  ─►  S02  Transformers (self-attention)
            └─ MISSING: GPT story  ─►  S03  GPT-1→3 + alignment (11 papers)
                 └─ MISSING: tokens & sampling  ─►  S04  BPE / Temp / Top-K / Top-P
                      └─ MISSING: how to RUN a model  ─►  S05  Ollama · LM Studio · OpenRouter
                           └─ MISSING: better prompts  ─►  S06  DSPy · CoT · MIPROv2 · GEPA
                                └─ MISSING: one API for every model  ─►  S07  LangChain LCEL
                                     └─ S08 Recap (S00–S07 visual pass)
                                          └─ MISSING: AI that writes the code  ─►  S09  /goal · /loop
                                               └─ MISSING: YOUR documents  ─►  S10  RAG + memory + chatbots
                                                    └─ MISSING: loops, pause, teams  ─►  S11  LangGraph
                                                         └─ MISSING: files, skills, subagents  ─►  S12  Deep Agents
                                                              └─ MISSING: one product  ─►  S13  Dining Bot
                                                                   └─ MISSING: ship it — cloud, identity, real users  ─►  S14  Bedrock AgentCore + AWS
                                                                        └─ MISSING: the model still talks like a generic assistant  ─►  S15  Fine-Tuning (LoRA / QLoRA)
                                                                             └─ S16 Full-course recap (S00–S15; does not replace S08)
                                                                                  └─ MISSING: even a fine-tuned LLM writes text when you only needed a decision  ─►  S17  Decision Models (RLCD · Jev)
                                                                                       └─ MISSING: the skills still need a product a hiring manager can click  ─►  S18  Final projects  ← you are here

---

👨‍🏫 Your instructor

Mohamed Noordeen Alaudeen — Data Scientist at AWS Generative AI Innovation Center · Dubai

• 🏆 Emerging Global Leader in GenAI — Internet 2.0 Conference Award (2024)
• 👥 29,000+ LinkedIn followers · 1,000+ professionals mentored
• 📚 Packt author (Data Science Interview Questions) · IIM Lucknow · 10+ years in AI/ML

“Learn GenAI the way professionals actually use it — not just theory, but systems that get you hired.”

---

📁 How this repo is laid out

One numbered folder per weekend. Open that folder’s README first. Class slides are linked in the session table, next to the notebooks in the syllabus, and collected in one list.

S11 extra (separate repo, not in this clone): Medium article agent.

---

🗺️ Sessions shipped so far

---

📚 Full syllabus (every session)

Setup lives in each session README. If that weekend has slides, they are linked in the section (and again in the presentation index).

---

Pre-work — Python, math, neural nets

If S00 looks like a stretch, start here. About three hours of Python, vectors, and how a network learns.

Also: prereq/cheatsheet.md — keep it open during S00.

---

S00 — How Search Engines Work

MISSING after this: TF-IDF misses "car crash" when you search "automobile accident". That is why S01 exists.

Slides: GenAI intro · How search works · Claude Code leak

Browser: How Search Engines Work (~45 min, no API key)
Write-up: How I Taught 100 Students to Build Google’s Core Algorithm in 30 Minutes

No API key. Pure Python.

---

S01 — Text to Numbers

MISSING after this: "bank" (river) and "bank" (finance) share one vector. S02 adds context.

App: CineMatch — FastAPI + React, 5 embedders on 1,000 IMDB movies.

Assignments: Medium article on the 5 methods · Medium article cosine vs Euclidean · product recommender (Amazon descriptions).

Slides: Text to Numbers
Browser: Cosine Similarity & Movie Recommender
Write-up: Words Don’t Have Meaning. Sentences Do.

---

S02 — Transformer Architecture

MISSING after this: Training from scratch is expensive. S03 is how that architecture became GPT and then ChatGPT.

Paper: Attention_Is_All_You_Need.pdf
Slides: Transformer architecture

Do these in the browser before the notebook:

Assets: SelfAttentionFull.mp4 · attention GIF · architecture spreadsheet.

GPU note: building blocks run on CPU; the training loop wants Colab Pro / H100.

---

S03 — GPT Evolution & Alignment

M00 (foundations) ends here. S00–S03 take you from raw text to how modern assistants are trained and aligned.

Path: text prediction (GPT-1) → scale (GPT-2/3) → alignment (RLHF → CAI → DPO).

11 papers (PDF + beginner summary notebook each):

Summaries: paper_summaries/. PDFs: papers/.

Browser: BERT for text classification (after paper 4)

---

S04 — BPE, Temperature, Top-K, Top-P

MISSING from S03: How does the model turn words into IDs? How does it pick the next token?

Browser (do these before the Excel / notebooks): Tokens / BPE · Temperature, top-k, top-p · Context window budget
Write-up (BPE): One Word. Two Tokens. Here Is How BPE Builds Them.

Order that matters: temperature → top-K → top-P → sample. Getting this wrong is a common production bug.

---

S05 — Local LLMs & API Providers

MISSING from S04: One provider’s API is a lock-in. Run a model on your laptop, or switch cloud providers by changing one variable.

Demos: apps/multi_provider_race.py · apps/map_reduce_demo.py

Portfolio: Distill — FastAPI + React + Whisper classroom tool (contribute).

Browser: Your first LLM call · Local vs cloud

Slides: S05 session PDF

---

S06 — Prompt Optimisation (DSPy · MIPROv2 · GEPA)

MISSING from S05: We could call models. Prompts were still handwritten guesses. Treat them as typed, versioned, scored code.

Benchmark used throughout: ATIS airline-intent (26 classes).

Browser (before DSPy): Zero-shot / few-shot / CoT · JSON or bust

---

S07 — LangChain Fundamentals

MISSING from S06: Each notebook talked to one provider with custom glue. LangChain is one interface.

Providers in the same notebook: gpt-4o-mini · Claude Haiku · Gemini Flash-Lite · Ollama llama3.2 / qwen2.5.

Browser: Chatbots forget (why you must send the message list) · First LLM call

---

S08 — Recap (S00–S07)

Visual pass of the MISSING chain before agentic coding.

• Interactive presentation
• Full-text slides

---

S09 — Agentic Coding & Loop Engineering

MISSING from S07: You still typed every line. Spec the goal, verify with tests, loop until green.

Browser: ReAct with a calculator — thought → tool → observation → stop.

Demo (external): Bullish Stock Scanner V3. FastAPI + React, 308 tests, built through spec-driven loops.

---

S10 — RAG + Memory & Chatbots (M07 + M08 + M06)

MISSING from S09: Agents can write software. They still only know training data. RAG grounds answers in your PDFs, policies, and tickets.

Memory & chatbots (M06) live here too: short-term memory, summarisation, long-term Store, condense-question, streaming, guardrails, and HITL are in S10f + notebooks 13–14.

Start page: 10_RAG/README.md · notebook index: 10_RAG/notebooks/README.md

Browser: Tiny RAG · Chunking · Hybrid + RRF · Citations / refusals · RAG injection · Chat memory · Production challenges
Write-up (Tiny RAG): I Asked ChatGPT About Our Gym’s Refund Policy. It Invented One.

S10a  Why RAG
S10b  Chunking (LangChain) + same ideas in LlamaIndex
S10c  Embeddings → FAISS / Chroma / Pinecone
S10d  BM25 → hybrid (RRF) → reranking
S10e  RAGAS + DeepEval
S10f  Production chatbots = Memory & Chatbots track
S10g  Retrieval showdown on one Pinecone index
        │
        ├── 13–14  Production chatbot (+ durable memory)   capstone
        ├── 15      Multimodal RAG                          extra
        ├── 16      MCP helpdesk                            extra · required for S11c
        ├── RAG Studio                                      FastAPI + React portfolio
        └── 9 student group datasets                        real-company briefs

Core notebooks (required)

Memory & Chatbots — what S10f / 13 / 14 actually teach (M06)

The chatbot module sits next to retrieval on purpose. A production bot is RAG plus memory.

S11 reuses condense(), trim_history(), the checkpointer, and Store. We don’t teach them again.

Capstones & extras

9 student group datasets (cohort project)

Every group ships a cited, refusal-aware support bot and 8 ablation tables (parse → chunk → embed → store → retrieve → fuse → rerank → generate). Spec: REQUIREMENTS_OVERVIEW.md · metrics: EVALUATION_METHODOLOGY.md.

---

S11 — LangGraph (stateful agents, M10)

MISSING from S10: ProductionRAGChatbot.chat() is a straight line. It cannot retry retrieval, rewrite a vague question, or pause for a human. LangGraph is that control flow.

Start page: 11_LangGraph/README.md · notebooks: notebooks/README.md

Browser: ReAct · HITL · MCP

S11a  Fundamentals & agents     required
S11b  Human-in-the-loop         required
S11c  Multi-agent orchestrator  required
  ├── S11d  Reasoning patterns  bonus (interview map)
  ├── S11e  SQL agent           bonus (Chinook)
  ├── capstone_agentic_rag/     optional self-correcting RAG
  └── medium-article-agent      optional FastAPI + React (separate repo)

Clone nursnaaz/medium-article-agent (Python 3.12+, Node 18+). Not bundled in this repo.

API you actually type: StateGraph · MessagesState · ToolNode · tools_condition · Command · interrupt() · MemorySaver · InMemoryStore · create_agent · create_supervisor.

Day 3 graph (the one you run):

top_supervisor
  ├── knowledge_team → rag_agent (search_knowledge_base) + search_agent (web)
  └── ops_team       → sql_agent (reads) + ticket_agent (writes → interrupt())

cd 11_LangGraph/multi_agent_orchestrator
python3 -m streamlit run app.py

Try: “What is our refund policy?” · “How many open tickets does Jane Doe have?” · “Add a note that we offered a refund” (then type yes or no).

---

S12 — LangChain vs LangGraph vs Deep Agents

MISSING from S11: LangChain talks. LangGraph is steps you wrote. Deep Agents is that graph with file tools, AGENT.md, SKILL.md, your tools, and helpers already attached — so “also add eggs” is not a new Python node.

Start page: 12_deepagents/README.md · notebook · slides

API: create_deep_agent · memory= · skills= · tools= · subagents= · write_file (built in).

---

S13 — Dining Bot (capstone)

MISSING from S12: a product. Dining Bot is one restaurant-manager chat: RAG over policies, read-only SQL over orders, two MCP servers (weather + charts), add menu item only after HITL, and multi-step planning via Deep Agents (AGENT.md + skills). The LLM proposes; code validates and executes.

Start page: 13_Project_Implementation/README.md · teach dining_bot.py (one file, SECTION 0→8) · run modular app.py + dining_bot/ · requirement (docx)

cd 13_Project_Implementation
python3.11 -m venv .venv && source .venv/bin/activate   # deepagents needs 3.11+
pip install -r requirements.txt
python build_db.py
streamlit run app.py
# optional CLI: python app.py --ask "Show me daily revenue for last week."

---

S14 — Bedrock AgentCore & Production AWS Deployment

MISSING from S13: Dining Bot runs on your laptop with one shared .env key. That is a demo, not a product. S14 is how the same kind of agent gets a managed Runtime, its own identity (Memory / Gateway / Identity), and a real deploy — Cognito login, a CDN URL, a container API, and infrastructure you can tear down and rebuild from code.

Start page: 14_Bedrock_AgentCore/README.md

Six self-contained labs, numbered 01. → 06.. Every Runtime lab starts from an empty AgentCore account — you create your own Memory, Gateway, and Identity. Never paste a classmate’s MEMORY_ID, Gateway credentials, or Runtime ARN.

S14 lab 04  Compare LangGraph vs Strands vs CrewAI, locally, in one notebook   ← start here (no AWS)
S14 lab 01  Same ideas on AgentCore, framework = LangGraph (+ Harness tools)
S14 lab 02  Same ideas on AgentCore, framework = Strands   (Support Copilot)
S14 lab 03  Same ideas on AgentCore, framework = CrewAI    (Competitor Brief)
S14 lab 05  Ship it: Strands agent + Cognito + React + FastAPI + CDK — one staged CDK app
S14 lab 06  Same ship-it stack, taught as 10 cumulative folders instead of stage flags

Recommended order: 04 (compare locally, free) → 01 / 02 / 03 in any order, each is “from scratch” → 05 or 06 (needs lab 02’s Runtime ARN first — pick the staged CDK app or the 10 folders, or do both).

Labs 01–03 — one agent, three frameworks, same AgentCore arc

Each of these three is genuinely self-contained — you can do just one if that’s the framework you care about. The shape repeats on purpose: Memory (recall across turns) → Gateway (MCP tools behind Cognito M2M auth) → Identity (OAuth provider) → Runtime (agentcore deploy), so the second and third lab teach faster than the first.

Regions (the #1 support question in this session):

Always pass --env AWS_REGION=us-east-1 on agentcore deploy — the Runtime does not read your local .env.

Lab 04 — compare the three frameworks locally (no AWS required)

04.framework-power-agents/ — one notebook, powerful_agents_comparison.ipynb, builds the same capable agent three ways (tools, RAG, optional MCP/Gateway, optional Memory) so you see the architectural difference before paying for AWS:

cd 14_Bedrock_AgentCore/04.framework-power-agents
uv sync --python 3.11
cp .env.example .env   # set OPENAI_API_KEY
source .venv/bin/activate
jupyter notebook powerful_agents_comparison.ipynb

Labs 05 & 06 — ship the full stack on your own AWS account

Both labs deploy the same production stack — they’re two teaching paths to one destination, not two different apps. Pick one, or do both.

05.agentcore-production-deploy/ — takes the Strands Support Copilot from lab 02 and puts a real product in front of it: Cognito username/password login → CloudFront (one HTTPS URL) → S3 (React chat UI) + App Runner (FastAPI) → IAM-invoked AgentCore Runtime. No AWS keys ever reach the browser.

Browser → Cognito login → ID token (JWT)
   ↓
CloudFront (one HTTPS URL)
   ├─ /* , /config.json  → S3 (React login + chat)
   └─ /api/* , /health   → App Runner (FastAPI)
                                │  verifies JWT, then uses an IAM role — no keys — to
                                ▼  call bedrock-agentcore:InvokeAgentRuntime
                          AgentCore Runtime (Strands Support Copilot, from lab 02)

Both walk the same 10 checkpoints: empty stack → Cognito pool → SPA app client → demo user → private S3 → CloudFront → React login (chat locked) → FastAPI /health → JWT-locked /api/me → full AgentCore chat.

First CDK deploy is usually 10–20 minutes (one-time account bootstrap). App Runner is the main always-on cost — lab 05’s Cleanup section tears it down; cdk destroy does not delete the AgentCore Runtime/Memory/Gateway/Guardrail from lab 02, so clean those up separately when you’re done.

What we learned building this session: the multi-framework labs (01–03) share almost all of their setup — Memory, Gateway, and Identity are identical concepts across LangGraph, Strands, and CrewAI, only the agent code differs. So labs 05/06 deliberately reuse lab 02’s flagship agent rather than building a fourth one — the point of S14 is deployment shape, not a new agent. Lab 06 exists as its own folder (not buried inside lab 05) because “one big CDK diff” hides which resource unlocked which capability; ten small, runnable, standalone snapshots make that visible.

---

S15 — LLM Fine-Tuning (PEFT · LoRA · QLoRA)

MISSING from S14: the agent is in the cloud, but it still talks like a generic model. Fine-tuning changes behavior (tone, format, domain phrasing). RAG still owns facts that change and need citations. This session teaches the decision, then the math that makes LoRA / QLoRA practical on one GPU.

Start page: 15_LLM_Fine_Tuning/README.md

---

S17 — Decision Models (RLCD · Jev)

MISSING from S15 / S16: a fine-tuned LLM still writes text, and its confidence is still something you cannot act on. Decision models return a typed pick plus a calibrated probability. You automate the sure ones and send the rest to a person.

Start page: 17_Decision_Models/README.md

Copy 17_Decision_Models/Jev/.env.example to Jev/.env and set JEV_KEY. The comparison also needs OPENAI_API_KEY in 17_Decision_Models/.env.

---

S18 — Final projects

MISSING from S17: a decision model can tell you which call is safe. It does not modernize the .NET Framework desk the company still runs.

Start page: 18_Final_Project/README.md

Apex Counter ships the legacy Web Forms app. Students convert a copy with AWS Transform, then run the ground-truth, evaluate, and repair prompts in that project’s README.

---

🏗️ Projects you can ship

What they look like

Live captures of the apps in this repo (same style as the Medium article agent README). Click a screenshot to open that project.

Dining Bot — RAG + SQL + HITL + MCP + Deep Agents planning for a restaurant manager.

Helpdesk Orchestrator — supervisor routes RAG, web, SQL, and ticket writes that pause for you.

Self-Correcting Agentic RAG — grade, rewrite, RAGAS, escalate.

Medium Article Agent — 23-node editorial graph (separate repo).

RAG Studio — ingest, swap strategies, chat, evaluate.

Production RAG chatbot — hybrid retrieval + rerank + cited answers.

RAG chatbot + memory — same engine, durable conversation memory.

Distill — transcript in, concept map + adaptive quiz out.

CineMatch — five embedders on the same 1,000 movies.

Holmes GPT — a GPT you trained yourself, streaming one token at a time.

---

🎮 Classroom presentations

GitHub’s file view shows HTML source, not slides. Use the links below, or the full index.

S00–S11 browser labs (no API key) live on nursnaaz.github.io. Same pairing as that site’s README. Medium write-ups are listed only when the article is live.

S04, S06, S07, S09, S10e, and S10g are notebook- or guide-led. S11c is the Streamlit orchestrator. S13 is the Dining Bot. S14 is lab-README-led (no browser tutorial). S15 has the fine-tuning deck + Unsloth notebook. Use 16_Full_Course_Recap/ for an end-of-arc pass through S00–S15. S17 has the decision-models deck.

---

🛠️ Tech stack (when each tool first appears)

---

🧭 Where the 23-module syllabus stands

S14 goes beyond the original 23 modules. Bedrock AgentCore hosting, Cognito authentication, and CDK infrastructure-as-code were not in the M00–M22 plan — they came out of a straightforward question after S13: the agent works, now how do I actually ship it? Treat S14 as the course’s first cloud-deployment module, layered on top of the syllabus rather than replacing a slot in it. S15 fills M19 (fine-tuning) that the earlier map still listed as ahead.

---

---

🤝 Contributing

Cohort questions go to WhatsApp, not Discord. For typos, broken links, or notebook bugs, see CONTRIBUTING.md.

---

📄 License

Source code and course materials in this repository are released under the MIT License.

---