9. Practice & Resources

Quick Reference Guide

Task	Library / Tool	Command
Run local LLM	Ollama	`ollama pull llama3.2`
Chat completions	openai / litellm	`litellm --model gpt-4o`
Vector search	ChromaDB	`Chroma.from_texts(chunks, embeddings)`
RAG pipeline	LangChain	`RetrievalQA.from_chain_type(...)`
Fine-tuning	TRL + PEFT	`SFTTrainer(model, dataset)`
QLoRA	bitsandbytes	`load_in_4bit=True`
UI prototyping	Gradio	`gr.ChatInterface(...).launch()`
Serverless deploy	Modal	`modal deploy app.py`

Useful Resources

Ollama — Local LLM runner
HuggingFace — Models, datasets, Spaces
LangChain — RAG and agent framework
Gradio — ML demo interfaces
Modal — Serverless GPU compute
TRL (Transformer Reinforcement Learning)

8-Week Study Plan

Week	Topic	Activities
1	LLM Product	Ollama, LLM types, streaming, brochure generator
2	Multi-Modal Chatbot	Gradio, tool calling, DALL-E, TTS, SQLite
3	HuggingFace	Pipelines, tokenizers, quantization, Whisper
4	LLM Showdown	Model comparison, benchmarks, code gen
5	RAG	LangChain, ChromaDB, evaluations, advanced RAG
6	ML/DL/Fine-Tuning	XGBoost, PyTorch, SFT, monitoring
7	Open-Source Fine-Tuning	QLoRA, adapters, W&B, evaluation
8	Multi-Agent Systems	Planning agents, Modal, deal scanner capstone

Certificate Path

AI-900: Azure AI Fundamentals
AI-102: Azure AI Engineer Associate
HuggingFace Course — Free NLP and diffusion model training

✏️ Final Capstone: Combine everything into a production-ready system: (1) fine-tune a small model on your domain data using QLoRA, (2) deploy it with Modal, (3) build a RAG pipeline around it with ChromaDB, (4) add a Gradio chat interface, (5) implement multi-agent orchestration with planning and review, and (6) monitor with W&B logging.