Build, measure and ship LLM applications with LangChain, LangGraph and LangSmith. Every module ends in a working, tested lab, and certification rests on a timed debugging exam and a capstone backed by evaluation evidence, not on attendance.
Students and working developers sit the same exam and are marked on the same capstone rubric, so the certificate means the same thing on every résumé.
Each three-hour module is roughly 60% hands-on. The amber notes are things we measured while building the course, and they are taught as findings.
A model call as a typed, testable function, with the provider chosen by configuration.
Lab: typed GST invoice extractorThe tool-calling loop, errors returned as text, and call budgets with middleware. Small local models often write tool calls as text instead of making them; choosing the model is part of the engineering.
Lab: campus assistant with a call budgetAnswers cite the clause they came from and refuse when they cannot, using a measured similarity floor and a citation check.
Lab: college-handbook Q&AGolden sets, hit rate, MRR, BM25 and reciprocal rank fusion. On our paraphrased questions, hybrid search did not beat vector search, and the lab teaches how to find that out before shipping.
Lab: retrieval scorecard and ship decisionBusiness rules in edges rather than prompts, and conversations that survive a restart.
Lab: student-support ticket graphA supervisor coordinates specialist agents, and nothing irreversible happens without a named person's approval. One model looped until the step limit stopped it, so the lab enforces progress in code.
Lab: approval-gated parent emailsExact evaluators gate the build, and an LLM judge is calibrated against human labels before anyone trusts it.
Lab: eval suite and CI gateA FastAPI service with a threat model: injection screening tested both ways, Aadhaar/PAN/phone redaction for DPDP-safe logs, rate limits, and a cost per request in rupees.
Lab: production API and containerBoth parts must be passed. The rubric is shared with learners in advance.
Find and fix three bugs in a working LangChain application, one each in a chain, a retriever and a graph, using the tests as your guide. Learners sitting next to each other get different bugs.
A working application with an evaluation set of 25 or more items, a calibrated judge and a written threat model. The panel asks what fails first. Unreproducible numbers or personal data in the repo fail outright.