AI Study

Generative AI fundamentals
Learn AI vs ML vs deep learning vs Generative AI.
Understand LLMs, training, inference, transformers, attention, context windows and hallucinations.
Understand where generative AI fits in full-stack applications.
Draw an LLM application architecture.

LLM concepts and tokenization
Tokens, token limits, context windows and temperature.
Input/output tokens, latency, cost and model selection.
System, developer and user instructions; conversation history.
Compare responses at different temperatures and prompt lengths.

First LLM API using Node.js
Set up a TypeScript project, environment variables and API key.
Make a request using the official JavaScript SDK.
Handle API errors, timeouts and token usage. Deliverable: A command-line AI assistant.
https://developers.openai.com/api/docs/quickstart

Prompt engineering — Part 1
Zero-shot and few-shot prompting.
Role, context, task, constraints and output format.
Clear instructions and examples.
Create prompts for summarization, translation and code review.

Prompt engineering — Part 2
Structured outputs, JSON schemas and Zod validation.
Prompt templates, output consistency and refusal handling.
Test prompts against a fixed set of questions. Deliverable: A resume analyzer that returns validated JSON.
https://developers.openai.com/api/docs/guides/structured-outputs?api-mode=responses

Embeddings and semantic similarity
Understand embeddings as numerical representations of text.
Learn cosine similarity, dimensions and semantic search.
Compare keyword search with meaning-based search.
Practice: Find similar FAQs using embeddings.

Function calling and weekly revision
Understand tools, function schemas and tool results.
Let an LLM call a safe calculator or product lookup function.
Revise all concepts and commit your code to GitHub. Milestone: An LLM can return structured answers and invoke tools.

RAG architecture
Learn why RAG is needed, how it differs from fine-tuning, and the ingestion vs query-time workflow.
Practice: Design the architecture for a company-document chatbot.

Document ingestion and chunking
Parse documents, split text into chunks, attach source metadata and preserve document IDs.
Deliverable: A script that ingests a folder of technical documents.

Vector databases
Learn vectors, indexing, similarity metrics, metadata filtering and persistence.
Start with PostgreSQL + pgvector; explore a dedicated vector database later.
Practice: Store document chunks and their embeddings in the database.

Semantic retrieval
Implement similarity search, top-k retrieval, metadata filters and relevant-document scoring.
Deliverable: A search API that returns the best matching chunks.

Build RAG with LangChain.js
Learn document loaders, text splitters, embeddings, vector stores, retrievers and prompt templates.
Deliverable: Ask questions about your own technical documentation.
https://reference.langchain.com/javascript/langchain

Improve RAG quality
Learn chunk size and overlap, hybrid search, reranking, retrieval thresholds, citation generation and "I don't know" responses.
Practice: Test questions that cannot be answered from your documents.

Build the first complete RAG API
Create Express endpoints for document ingestion, question answering and source retrieval.
Milestone: A working document Q&A backend with citations.

AI agent fundamentals
Agent loop: observe, decide, act and evaluate.
Tools, tool selection, observations and stopping conditions.
Understand ReAct-style agents and when a fixed workflow is better than an autonomous agent.
Practice: Design an agent that can search internal product documentation.

LangChain.js fundamentals
Learn chat models, prompt templates, output parsers, runnables, retrievers and tool definitions.
Deliverable: A reusable model-and-retriever service.
https://reference.langchain.com/javascript/langchain

Tool calling and agent implementation
Build tools such as searchKnowledgeBase, getProductDetails and calculateTotal.
Validate tool arguments with Zod and return typed results.
Deliverable: An agent that chooses between a calculator and RAG search.

Agent memory and conversation history
Learn short-term conversational state, persistent checkpoints, summarization and retrieval of previous context.
Practice: Maintain conversation context across multiple questions.

LangGraph fundamentals
Learn graph state, nodes, edges, conditional edges, loops, checkpointing and termination conditions.
Deliverable: A graph that classifies a question, retrieves information and drafts an answer.
https://docs.langchain.com/oss/javascript/langgraph/thinking-in-langgraph

Advanced agent workflows
Conditional routing and retry logic.
Human approval before sensitive actions.
Persistence, recovery and agentic RAG.
Separate planning, retrieval and answer-generation steps. Practice: Require human approval before an agent executes a mock refund.

Agent evaluation and safety
Test tool-selection accuracy, invalid tool arguments, loops, timeouts and incorrect answers. Learn traces, execution logs and regression tests.
Milestone: A stateful LangGraph agent with tools, error handling and approval.
https://github.com/langchain-ai/langgraphjs

Set up the project
Create React/Vite and Node/Express TypeScript projects.
Configure environment variables, API routes, error handling and GitHub.

Build the chat interface
Create the chat screen, message history, loading states, error states and responsive UI.

Implement secure LLM APIs
Build backend-only model calls, request validation, timeouts, rate limits and streaming responses.

Add document ingestion
Accept appropriate document formats, extract text, chunk documents, create embeddings and save chunks with metadata.

Connect vector retrieval
Implement pgvector similarity search, top-k retrieval, metadata filtering and document-level access restrictions.

Implement RAG answers
Pass retrieved chunks to the model, return source references and handle insufficient evidence gracefully.

Add a LangGraph agent
Create workflow nodes for intent classification, knowledge retrieval, tool execution and response generation.

Testing and production hardening
Test APIs and RAG quality. Add authentication, permission checks, prompt-injection defenses, logging and cost/latency tracking.

Deploy and prepare interviews
Deploy frontend, backend and database. Write README and architecture diagram, record a demo and prepare project-based interview answers.

Study Material
LLM APIs and Generative AI
https://developers.openai.com/api/docs/quickstart
https://developers.openai.com/api/docs/guides/structured-outputs?api-mode=responses
Learn model integration, structured results and tool calling.

LangChain.js
https://docs.langchain.com/oss/javascript/langchain/overview
https://reference.langchain.com/javascript/langchain
Focus on model integrations, retrievers, tools, prompts and RAG.

LangGraph.js
https://docs.langchain.com/oss/javascript/langgraph/overview
https://github.com/langchain-ai/langgraphjs
Focus on state, nodes, conditional routing, checkpointing and approval workflows

Vector databases and RAG
https://supabase.com/docs/guides/ai
https://supabase.com/docs/guides/database/extensions/pgvector
Learn embeddings, vector similarity, storage and retrieval.