Building LLM Applications : Agents, Memory, RAG (Hands-on)

Workshop equips software teams with practical skills to build tool-using AI agents that retrieve knowledge, call APIs, and follow rules reliably. Participants will design prompts, enforce structured outputs, implement RAG with a vector store, orchestrate workflows with LangChain, serve agents via FastAPI, run locally with Ollama, and improve quality using LangSmith tracing and evaluations.

• Build a solid understanding of LLM fundamentals including tokens, context windows, temperature, and failure modes to make better engineering decisions.
• Equip participants to write reliable, structured prompts and run AI models locally using Ollama for privacy-sensitive or cost-conscious environments.
• Design and deploy agent services with real tool calling using FastAPI and LangChain.
• Implement RAG pipelines that ground agent responses in real documents using ChromaDB for accurate, citation-backed outputs.
• Close the quality loop by tracing, evaluating, and improving agent performance using LangSmith.

1. Module 1

   Foundations + AI-Ready Engineering Patterns

   LLM core concepts, prompt reliability patterns, and structured JSON output. Participants build a prompt pack applying role, rules, examples, and failure handling to real developer scenarios.

2. Module 2

   Local vs Cloud Models + Running with Ollama

   Choosing between local and hosted models across privacy, cost, and speed. Covers Ollama setup, context strategy, and prompting best practices for smaller local models.

3. Module 3

   Tool Calling + Agent Service with FastAPI

   What makes an agent work — model, tools, memory, and context. Participants build a live FastAPI agent endpoint with two connected tools using LangChain.

4. Module 4

   RAG Grounding with LangChain Document Loaders + ChromaDB

   Chunking, embeddings, retrieval, and citations explained and applied. Participants build a grounded Q&A pipeline from PDF loader through ChromaDB to a cited answer.

5. Module 5

   Agent Orchestration with LangChain (Multi-Step Workflows)

   Chains, routers, planners, and tool loops for multi-step workflows. Covers session vs durable memory and culminates in a full retrieve → decide → act → format pipeline.

6. Module 6

   Tracing + Evaluations with LangSmith

   Debugging agent failures through full run traces. Participants add LangSmith to existing workflows, build a golden eval set, and run pass/fail regression tests.