AI Engineering
Learn LLMs, RAG, agents, secure tools, and production AI through practical insurance examples.
A tutorial's imports fail. What do you do?
I identify its intended version, consult the current official docs, and separate the concept from the old API. I pin compatible dependencies and avoid mixing legacy chain APIs with newer agent construction patterns.
LangGraph
Beginner — What are nodes, edges, and state?
What are nodes, edges, and state?
Nodes are steps that perform work, edges choose what runs next, and state carries the information needed by those steps. Nodes can be ordinary code or model calls, so the graph can be deterministic or hybrid.
What does a checkpointer do?
It saves execution progress for later continuation or recovery. It is not automatically long-term user memory or a transaction coordinating all external effects. Production needs durable storage and authorized thread ownership.
What makes a graph production-ready?
Explicit state contracts, safe concurrent updates, durable checkpoints, version migrations, authenticated resume, bounded retries/deadlines, idempotent side effects, and redacted tracing. I also define how pending runs survive deployment and how uncertain…
Resuming an approval sends an email twice. Why?
The interrupted node may restart, so a send before the interrupt can execute again. I move sending behind validated approval into an idempotent business operation and reconcile the send key/receipt on retries.
AI agents
Beginner — Agent versus chatbot?
Agent versus chatbot?
A chatbot is a conversational interface. An agent uses a model to choose bounded actions and observes results toward a goal. A chatbot can use an agent, but a one-call conversation does not require one.
What is an agent loop?
The model proposes a step, the application validates and executes a permitted action, and the result becomes an observation. The loop repeats only while task progress and step/time/token limits allow it.
When are multiple agents justified?
When distinct scopes or specialties show measurable benefit over one agent or a fixed workflow. I require typed responsibilities, limited delegation, shared budgets, consistency checks, and outcome evaluation—not cosmetic role-play.
An agent keeps repeating search. How do you fix it?
I inspect whether the goal and completion conditions are clear and whether tool results are usable. I add repeated-query/progress detection and strict budgets, then clarify or escalate when evidence remains insufficient rather than searching indefinitely.
MCP
Beginner — What problem does MCP solve?
What problem does MCP solve?
It standardizes how AI applications connect to and discover external tools and context. Multiple approved hosts can reuse the same server-side capability contracts rather than each building a custom integration.
MCP versus function calling?
Function calling lets a model propose a structured operation. MCP provides an integration protocol for discovering/invoking capabilities. The host can turn a model proposal into an MCP request that calls an existing REST service underneath.
How do you secure an MCP server?
I authenticate requests, verify token properties, authorize each tool and object, restrict downstream credentials/egress, validate schemas/results, review tool definitions, and audit actions. Connection permission does not equal customer-record permission.
A new discovered tool requests secret access. What happens?
The host should not automatically expose or execute it. Tool descriptions are untrusted unless reviewed; allowlists and execution policy block capabilities outside the task. I investigate the server/definition change and preserve an audit record.
Microsoft Agent Framework
Beginner — What is MAF?
What is MAF?
Microsoft Agent Framework provides agents and explicit workflows. An agent handles model-driven interaction and tools; a workflow connects agents and functions through controlled execution paths. It builds on ideas from AutoGen and Semantic Kernel.
Difficulty is an editorial estimate. Completion means you chose to mark a question as understood.