- Learn
- Artificial Intelligence
- Capstone Projects
AI-M15
Capstone Projects
Six guided builds that put the whole track to work: a cited RAG app, an MCP tool server and agent, a text-to-SQL analyst, a code reviewer, a voice assistant and an eval harness.
Reading about retrieval, agents and evals is not the same as making them work together. Each capstone in this module is a guided build: a clear goal, the architecture, a sequence of checkpoints that each end with something you can run and verify, and a final checklist of what "done" means.
Every capstone comes with a small, runnable core in plain Python that uses a fake model, so you can see the whole system work end to end on your laptop before you spend anything on API calls. Then you swap in a real model, embeddings or speech services and measure the result. You will build a chat-with-your-documents app with citations, an MCP server with two tools behind an agent loop, a natural-language analyst that writes safe SQL, an AI code reviewer judged against known bugs, a voice assistant with a latency budget, and finally the eval harness that keeps all of them honest.
Lessons
- AI-15.01 Capstone: chat with your documents (RAG) In this capstone you build a chat-with-your-documents app: ingest a document set, chunk and embed it, store vectors, retrieve with hybrid search and reranking, generate answers with citations and refusal when unsupported, and measure quality with a small eval set. Intermediate 60 min
- AI-15.02 Capstone: a tool-using agent with MCP In this capstone you write an MCP server exposing two tools, connect it to an agent loop, and handle tool errors, step limits and confirmation for risky actions, then evaluate whether the agent completes a set of realistic tasks. Advanced 60 min
- AI-15.03 Capstone: text-to-SQL analyst In this capstone you build a natural-language-to-SQL analyst: retrieve relevant tables and columns, generate SQL, validate it, run it read-only with limits, and explain results, with tests for both correct queries and safe failure. Advanced 45 min
- AI-15.04 Capstone: AI code reviewer In this capstone you build an AI code reviewer that reads a diff, produces structured findings with file, line, severity and explanation, and is evaluated with an LLM judge and a set of known bugs to measure precision and recall. Advanced 45 min
- AI-15.05 Capstone: voice assistant In this capstone you build a voice assistant that streams speech to text, generates replies with a language model and speaks them back, measuring end-to-end latency for each stage and optimizing until conversation feels natural. Advanced 45 min
- AI-15.06 Capstone: build your own eval harness In this capstone you build an evaluation harness: load datasets, run a system under test, grade outputs with code checks and LLM judges, produce reports and track results over time so regressions are caught before release. Intermediate 40 min