AI Engineering with LLMs: Advanced

Twenty-two pages for after the beginner tutorial, built on the same EvidenceDesk kit. You improve retrieval with hybrid search, rank fusion and reranking, measured against a labelled corpus; build a local MCP server and client with validated inputs and bounded results; and add a controlled agent with step limits, explicit state and human approval for consequential actions. You then build evaluation sets with holdouts and release gates, and test an LLM judge for bias before trusting it.

Later parts cover fine-tuning as a measured alternative to better context, with LoRA and PEFT explained through a tiny numeric toy and a dry-run experiment record, not real training. Then LLMOps (structured logs, tracing, cost and monitoring), threat modelling, injection testing and guardrails, deployment with streaming, scaling and rollback, and permission-aware ingestion and data freshness. The capstone, KnowledgeOps, assembles a release candidate with evidence and ends with an honest no-go decision. A graded practice lab and the capstone grader finish the tutorial.

Everything runs offline with fake models, toy embeddings and mocked services, so no paid API, GPU or internet is needed, and no real GPU training is run. Docker files are shown and checked statically, not run. The tutorial is educational: facts were checked on 12 October 2026, and model names, prices and SDK details change. The Bangla text was written by AI. Every page is available in English and Bangla.