Applied Agentic AI Engineering for Banking Professionals

Builds essential Python skills for banking AI applications. Covers data structures, file I/O, error handling, and CSV processing. All labs use banking scenarios including reading bank statements, validating transactions, flagging suspicious values, and exporting structured JSON summaries. Produces the BankDataProcessor capstone seed reused throughout the course.

Teaches reliable calling of GPT-4o and Claude APIs. Covers system/user/assistant roles, temperature, token counting, cost estimation, zero-shot/few-shot prompting, Chain-of-Thought reasoning, CO-STAR and RAISE frameworks. Produces the BankingLLMClient capstone seed—a unified interface supporting streaming, structured output, async batch processing, and per-call cost logging.

Builds retrieval systems enabling AI to answer questions using bank regulatory documents, policies, and FAQs with citations. Covers LCEL pipe operator, conversation memory (buffer, summary, entity), document loaders (PDF, CSV, web), chunking strategies, embeddings, cosine similarity, MMR/multi-query retrievers, source citation, RAGAS evaluation, and HyDE for follow-up questions. Produces BankBot360 Compliance Assistant with conversational RAG and source citations.

Migrates from in-memory vector store to persistent databases—first locally with Chroma, then to production Pinecone with namespace isolation. Covers CRUD operations, metadata filtering, batch upsert, and LangChain integration. Produces a VectorStore factory layer that switches between Chroma and Pinecone via environment variable, enabling Phase 3 RAG pipelines to run unchanged over the abstracted layer.

Moves from linear chains to decision-making agents. Covers ReAct pattern (Reason, Act, Observe), StateGraph, TypedDict state schema, conditional edges, checkpointing (MemorySaver, PostgresSaver), @tool decorator with Pydantic schemas, six banking tools, ToolNode, and human-in-the-loop patterns (interrupt, resume, officer approval workflows). Produces BankBot360 Fraud Detection Agent—a LangGraph ReAct agent with seven tools, human-in-the-loop interrupt, LangSmith tracing, and PostgreSQL checkpointing.

Covers the N×M integration problem MCP solves, three-layer architecture (Host, Client, Server), primitives (Tools, Resources, Prompts), and transport options (stdio vs SSE). Builds a production MCP server using FastMCP SDK exposing six banking tools, three resources (regulatory documents by URI), and two prompt templates. Produces complete BankBot360 MCP infrastructure validated in MCP Inspector, with SSE transport ready for production and LangGraph fraud agent decoupled via MCP.

Builds a team of five specialist AI agents collaborating to complete the KYC/AML pipeline. Covers Agent (role, goal, backstory, tools), Task (description, expected_output, context), Crew (Process.sequential, Process.hierarchical), backstory engineering, hierarchical crews with manager LLM delegation, and MCP tool integration. Produces BankBot360 KYC/AML Pipeline—a five-agent hierarchical crew (Document Collector, Sanctions/PEP Screener, AML Analyst, Risk Scorer, Compliance Report Writer) embedded as a node inside LangGraph OnboardingGraph.

Exposes every BankBot360 component as a production web service with audit trail for banking AI governance. Covers GET/POST endpoints, Pydantic request/response schemas, four AI routers (/chat, /compliance, /fraud, /kyc), WebSocket streaming, MCP SSE route alongside REST, SQLAlchemy async with four database tables (ChatMessage, AgentRun, MCPToolCall, KYCRecord), JWT authentication, slowapi rate limiting, CORS, and pytest with 70%+ coverage. Produces BankBot360 FastAPI backend with full audit trail and 70%+ test coverage.

Integrates everything into a usable interface, containerised stack, and live team presentations. Covers Streamlit three-page dashboard (Compliance Chat with source citations, Fraud Investigation with agent reasoning trace, KYC Review with five-agent progress), Dockerfile for API and UI, docker-compose.yml for full stack, and GitHub Actions CI workflow. Culminates in team capstone presentations demonstrating the complete BankBot360 system.