Patternwise

AI track

AI Engineering

From gradient descent to production LLM systems.

A path from machine-learning fundamentals through LLM applications to running AI systems in production. Each module builds only on what came before, and every lesson ends with the questions interviewers actually ask.

3stages
19modules
49lessons
10projects
10hours
Start the first lesson
0/49 lessons
Stage 1

Foundations

How models learn, how transformers work, and how LLMs are trained and sampled — the mental model everything else rests on.

Module 014 lessons · 44 min

Machine Learning Basics

Explain supervised vs unsupervised learning, split data correctly, spot overfitting, and describe how gradient descent minimizes a loss.

  1. ✓What Machine Learning Actually IsLearning a function from examples: supervised, unsupervised, and self-supervised learning.10m
  2. ✓Train, Validation, and Test SplitsWhy you hold data back, how to split it, and the leakage traps that fake good results.10m
  3. ✓Overfitting and RegularizationMemorizing vs generalizing, the bias–variance trade-off, and the tools that keep models honest.12m
  4. ✓Loss Functions and Gradient DescentHow a model measures its mistakes and walks downhill to make fewer of them.12m
Module 022 lessons · 24 min

Neural Networks

Describe a neural network as stacked linear layers plus non-linearities, and explain backpropagation without hand-waving.

  1. ✓Neurons, Layers, and ActivationsWhat a layer computes and why non-linear activations make depth worthwhile.12m
  2. ✓Backpropagation IntuitionThe chain rule as blame assignment: how every weight learns its share of the error.12m
Module 032 lessons · 24 min

Deep Learning Building Blocks

Recognize the components that make deep networks trainable and read a training loop with confidence.

  1. ✓Normalization, Dropout, and ResidualsThe three tricks that let very deep networks train stably.12m
  2. ✓Optimizers and Training LoopsSGD, momentum, Adam, learning-rate schedules, and the anatomy of a training step.12m
Module 042 lessons · 20 min

Tokens and Embeddings

Explain how text becomes tokens and tokens become vectors, and why that matters for cost, limits, and meaning.

  1. ✓TokenizationSubword tokenizers like BPE, and why token counts drive cost and odd model behavior.10m
  2. ✓Embeddings as MeaningTurning tokens and texts into vectors where distance tracks similarity.10m
Module 052 lessons · 28 min

Attention and the Transformer

Walk through self-attention with queries, keys, and values, and sketch a decoder-only transformer block end to end.

  1. ✓The Attention MechanismQueries, keys, and values: how each token decides what to look at.14m
  2. ✓The Transformer ArchitectureMulti-head attention, feed-forward layers, positions, and causal masking in one block.14m
Module 063 lessons · 36 min

How LLMs Are Trained

Explain the pretraining → instruction tuning → preference tuning pipeline and what each stage adds.

  1. ✓Pretraining: Next-Token PredictionHow predicting the next token on huge corpora produces general capability.12m
  2. ✓Supervised Fine-TuningTurning a text-completer into an instruction-follower with curated demonstrations.10m
  3. ✓RLHF and Preference TuningReward models, PPO, DPO, and how human preferences shape model behavior.14m
Module 072 lessons · 22 min

Sampling and Decoding

Choose decoding settings deliberately and use logprobs for confidence, classification, and debugging.

  1. ✓Temperature and Top-pHow the next token is actually chosen and which knobs change it.10m
  2. ✓Logprobs and Decoding StrategiesGreedy, beam, and constrained decoding, plus what logprobs can tell you.12m
Stage 2

Applications

Building with LLMs: prompting, retrieval, tools, agents, MCP, LangGraph, and knowing when to fine-tune.

Module 083 lessons · 36 min

Prompt Engineering

Write prompts that are structured, testable, and reliable, including ones that return machine-parseable output.

  1. ✓Prompt StructureSystem vs user messages, roles, delimiters, and writing instructions a model can follow.12m
  2. ✓Few-Shot and Chain-of-ThoughtTeaching by example and giving the model room to reason before answering.12m
  3. ✓Structured Output and JSONGetting reliable JSON with schemas, constrained decoding, validation, and repair.12m
Module 092 lessons · 27 min

Embeddings and Vector Search

Build semantic search and explain how ANN indexes like HNSW, IVF, and PQ trade recall for speed and memory.

  1. ✓Semantic Search with EmbeddingsEmbedding documents and queries, similarity metrics, and where keyword search still wins.12m
  2. ✓Vector Databases and ANN IndexesHNSW, IVF, and product quantization, and how to choose a vector store.15m
Module 103 lessons · 40 min

Retrieval-Augmented Generation

Design a RAG pipeline end to end: chunking, hybrid retrieval, reranking, query rewriting, and grounded answers with citations.

  1. ✓RAG Fundamentals and ChunkingThe ingest/retrieve/generate loop and how chunking choices make or break it.14m
  2. ✓Hybrid Search and RerankingCombining BM25 with vectors, fusing rankings, and reranking with cross-encoders.14m
  3. ✓Query Rewriting and CitationsFixing bad queries before retrieval and making answers verifiable after it.12m
Module 114 lessons · 50 min

Tool Calling and Agents

Implement a tool-calling loop, explain ReAct and planning patterns, and design agent memory and multi-agent setups.

  1. ✓Tool and Function CallingHow models request tool calls, how your code executes them, and how to design good tools.12m
  2. ✓ReAct and Planning AgentsReason–act–observe loops, plan-and-execute, and keeping agents from spinning.14m
  3. ✓Agent MemoryShort-term context, long-term stores, summarization, and what to remember.12m
  4. ✓Multi-Agent SystemsSupervisors, handoffs, and when splitting work across agents helps or hurts.12m
Module 122 lessons · 24 min

Model Context Protocol

Explain MCP's host/client/server model and build a server that exposes tools, resources, and prompts.

  1. ✓MCP ArchitectureHosts, clients, servers, transports, and why a standard protocol for context matters.12m
  2. ✓MCP Tools, Resources, and PromptsThe three server primitives, who controls each, and how to design them well.12m
Module 132 lessons · 28 min

LangGraph

Model an agent as a state graph with nodes, edges, and reducers, and add persistence and human approval steps.

  1. ✓State Graphs, Nodes, and EdgesTyped state, reducers, conditional edges, and cycles in LangGraph.14m
  2. ✓Checkpoints and Human-in-the-LoopPersisting graph state, resuming threads, interrupts, and time travel.14m
Module 142 lessons · 26 min

Fine-Tuning

Decide between prompting, RAG, and fine-tuning for a given problem, and explain how LoRA makes tuning cheap.

  1. ✓Fine-Tuning vs RAG vs PromptingA decision framework: knowledge problems, behavior problems, and cost.12m
  2. ✓LoRA and Parameter-Efficient Fine-TuningLow-rank adapters, QLoRA, and the practical workflow for tuning a small model.14m
Stage 3

Production

Shipping and running AI systems: evaluation, security, cost, observability, serving, and system design interviews.

Module 153 lessons · 38 min

Evaluation

Build offline eval suites with golden sets, use LLM-as-judge responsibly, and measure RAG quality component by component.

  1. ✓Offline Evals and Golden SetsTreating prompts like code: datasets, metrics, and regression gates in CI.14m
  2. ✓LLM-as-JudgeUsing a model to grade a model, its biases, and how to calibrate it against humans.12m
  3. ✓RAG Evaluation MetricsRetrieval recall, MRR, faithfulness, and answer relevance — measured separately.12m
Module 163 lessons · 36 min

Guardrails, Security, and Responsible AI

Threat-model an LLM app, defend against prompt injection and exfiltration, and build guardrails and responsible-AI practices into delivery.

  1. ✓Prompt Injection and Data ExfiltrationDirect and indirect injection, the lethal trifecta, and defenses that actually hold.14m
  2. ✓Guardrails, Jailbreaks, and PIIInput/output filters, jailbreak resistance, and keeping personal data out of logs and prompts.12m
  3. ✓Responsible AI in PracticeBias, transparency, human oversight, and turning principles into engineering tasks.10m
Module 173 lessons · 36 min

Cost, Latency, and Observability

Cut cost and latency with caching, streaming, batching, and routing, and trace every LLM call so you can debug production.

  1. ✓Caching and Prompt CachingExact, semantic, and provider-side prefix caching, and when each pays off.12m
  2. ✓Streaming, Batching, and Model RoutingPerceived vs real latency, batch APIs, and sending each request to the cheapest model that can handle it.12m
  3. ✓Observability and Tracing for LLM AppsSpans for every model and tool call, what to log, and closing the loop with evals.12m
Module 182 lessons · 26 min

Serving and Inference

Explain what happens inside an inference server — KV cache, continuous batching, quantization — and size a self-hosted deployment.

  1. ✓KV Cache and Continuous BatchingPrefill vs decode, why memory bandwidth rules, and how servers keep GPUs busy.14m
  2. ✓QuantizationShrinking weights to 8 or 4 bits and what you trade for the speed.12m
Module 193 lessons · 44 min

AI System Design

Run an AI system design interview with a repeatable framework and apply it to a support copilot and a RAG search system.

  1. ✓The AI System Design FrameworkA step-by-step structure for AI design interviews, from requirements to evals.14m
  2. ✓Design a Customer Support CopilotA worked design: retrieval, tools, guardrails, handoff, and metrics.15m
  3. ✓Design an Enterprise RAG Search SystemA worked design: ingestion at scale, permissions, hybrid retrieval, and freshness.15m

Then practise

Build Projects

Hands-on projects that turn lessons into portfolio pieces. Each one has milestones with acceptance criteria, a core code skeleton, and talking points for interviews.

Open build projects →