AI & Machine Learning

Beginner

No prior ML experience required. Build your intuition for how AI systems work and start calling real APIs.

1. The AI Landscape

   Understand what machine learning actually is, how it differs from classical programming, and where the different subfields (CV, NLP, RL) live.

   ~1 week

2. Python Foundations for ML

   NumPy, Pandas, and Matplotlib form the substrate everything else runs on. Understand array operations and data manipulation before touching any models.

   ~2 weeks

3. Your First ML Project

   Train a classifier end-to-end using scikit-learn. Focus on the train/validate/test split, overfitting intuition, and reading metrics honestly.

   ~1 week

4. Using LLMs via API

   Call the OpenAI and Anthropic APIs. Learn prompt construction, token economics, and when to use system prompts vs. few-shot examples.

   ~2 weeks

Intermediate

You can write Python and have used an LLM API. Now go deeper: understand the architecture, run models locally, and build RAG systems.

1. Deep Learning Architecture

   CNNs, RNNs, and the attention mechanism. Understand backpropagation intuitively, not just formulaically. This is the foundation for everything else.

   ~2 weeks

2. Transformers — How They Really Work

   Read and implement the "Attention Is All You Need" architecture. Understand tokenization, positional encoding, and why transformers scale the way they do.

   ~2 weeks

3. Running Local AI Models

   Set up Ollama on any hardware. Understand GGUF quantization, GPU offloading, and the tradeoffs between model size and quality.

   ~1 week

4. Building RAG Pipelines

   Retrieval-Augmented Generation solves hallucination by grounding the model in your data. Learn chunking strategies, embedding models, and vector database design.

   ~3 weeks

5. Fine-Tuning Open Models

   Adapt Llama, Mistral, or Gemma to your specific domain using LoRA and QLoRA. Use Unsloth for 5× faster fine-tuning on a single consumer GPU.

   ~2 weeks

Advanced

Production-grade systems: model serving at scale, multi-modal architectures, autonomous agents, and engineering for reliability.

1. MLOps & Model Serving

   Run vLLM, Text Generation Inference (TGI), or Triton Inference Server in production. Understand continuous batching, quantized serving, and latency vs. throughput tradeoffs.

   ~3 weeks

2. Multi-modal AI Systems

   Vision-language models (VLMs) like LLaVA, Qwen-VL, and Pixtral. Build systems that can reason over images, documents, and structured data simultaneously.

   ~3 weeks

3. Agents & Tool Use

   Build autonomous agents using ReAct, function calling, and the Model Context Protocol (MCP). Design for reliability: handle failures, loops, and unexpected model behavior.

   ~3 weeks

4. Production AI Architecture

   Evaluation pipelines, observability, cost management, and guardrails. Build systems you can actually trust in production.

   ~3 weeks

Do I need a GPU to get started with AI?

No. For learning, calling APIs (OpenAI, Anthropic, Groq) requires only an internet connection and a few dollars. For running local models, modern Apple Silicon Macs with unified memory are excellent — an M2 MacBook Pro can run Llama 3 8B comfortably. A dedicated GPU becomes important when fine-tuning at scale.

What is the difference between RAG and fine-tuning?

RAG (Retrieval-Augmented Generation) retrieves relevant documents at inference time and gives them to the model as context. Fine-tuning bakes knowledge into the model weights during training. RAG is better for frequently updated information and when you need source citations. Fine-tuning is better for adopting a specific style, tone, or domain vocabulary that should be deeply embedded.

Which local model should I start with?

For general purpose tasks, start with Llama 3.1 8B or Mistral 7B via Ollama — they run well on 16GB of RAM. For coding, Qwen2.5-Coder or DeepSeek-Coder are excellent. For anything requiring a larger context window, try Gemma 2 27B if your hardware supports it. Avoid 70B+ models unless you have a dedicated GPU setup.

Is Python mandatory for AI/ML work?

For most of the ML ecosystem, yes. PyTorch, Hugging Face, LangChain, and nearly all research code is Python-first. That said, inference APIs can be called from any language. Rust and Go are emerging for inference servers (llama.cpp has Go bindings), and JavaScript/TypeScript has Vercel AI SDK and LangChain.js for application-layer work.

What is quantization and why does it matter?

Quantization reduces a model's memory footprint by representing weights in lower precision (e.g., 4-bit instead of 16-bit floats). A 7B model at full FP16 requires ~14GB of VRAM; at Q4_K_M quantization it needs ~4GB with minimal quality loss. This is what makes running models on consumer hardware practical. GGUF is the most common format for quantized models you'll find on Hugging Face.

What is the Model Context Protocol (MCP)?

MCP is an open standard (developed by Anthropic) that defines how AI models can interface with external tools, data sources, and services in a standardized way. It's analogous to USB for AI agents — instead of each application implementing its own integration layer, tools expose an MCP server and any compatible model can use them. Adoption is growing fast across IDEs, desktop apps, and CLIs.