AI Engineering
Learn LLMs, RAG, agents, secure tools, and production AI through practical insurance examples.
Project 4 — Production RAG application
Requirements: Multi-tenant retrieval, incremental updates/deletions, policy versions, hybrid search, evaluated reranking, and source-aware caching.
Project 5 — Tool-calling agent
Requirements: Read claim status, lookup policy, retrieve instructions, ask for missing IDs, and escalate. Start read-only.
Project 6 — LangGraph agent
Requirements: Convert the assistant into an explicit state graph, add durable progress, clarification/approval pauses, and safe resume.
Project 7 — Multi-agent customer support
Requirements: Add specialists only for measurable policy/claim/document tasks, with clear coordinator ownership and escalation.
Project 8 — MCP server plus AI agent
Requirements: Expose reviewed customer lookup, order/policy lookup, ticket creation, and knowledge search to an approved host.
Project 9 — Deterministic plus agentic workflow
Requirements: Interpret input, prepare a draft, validate required fields, wait for review, execute once, and notify reliably.
Project 10 — Production insurance AI platform
Requirements: Combine policy RAG, claim status, guided intake, document processing, reviewed communication, staff escalation, and multi-tenant operations. Define supported tasks and explicit exclusions.
LLMs
Practice saying the answer clearly, then explain the insurance example and the trade-off. A strong answer identifies the authoritative data, safe failure, and measurement—not only the framework name.
What is an LLM good for?
An LLM processes language using learned patterns and supplied context. I use it to summarize, extract, classify, explain, and propose tool actions. Current claim status and business permissions come from services, not the model's memory.
Does temperature zero guarantee the same correct result?
No. It generally reduces sampling variation where supported, but it does not verify evidence or eliminate all runtime variation. I use schemas, authoritative data, bounded outputs, and evaluations to establish application reliability.
How do you choose a model?
I compare candidate models on representative tasks, including failure and adversarial cases. I measure correctness, supported tool/schema capabilities, latency, cost per successful task, and deployment/data restrictions, then version the selected…
The model invents a claim status. What do you change?
I remove any path that asks it to know account facts. An authorized tool returns canonical status; a template or model explains that result. I test missing claims, denied access, backend failure, and unsupported extra statements.
Prompt engineering
Beginner — What makes a good prompt?
What makes a good prompt?
A clear task, trusted constraints, explicitly separated input, expected output, and a defined response when information is missing. For policy answers, I require supporting evidence and forbid inventing unsupported policy conditions.
Zero-shot or few-shot?
I begin with clear zero-shot instructions. If errors concern category boundaries or formatting, I add a few representative reviewed examples and evaluate them on separate cases. Examples should cover ambiguity, not just easy successes.
How do you maintain prompts?
I treat prompts as versioned configuration with an owner, evaluation dataset, release gate, and trace tag. I investigate patterns in failures instead of accumulating conflicting instructions for individual anecdotes. Provider/model upgrades rerun the same…
Adding many examples improved a demo but slowed production. What next?
I measure which examples actually help and remove redundancy. I test a smaller prompt, provider-supported schemas, deterministic preprocessing, or a better model. I optimize latency and cost while preserving held-out quality.
RAG
Beginner — What is RAG?
Difficulty is an editorial estimate. Completion means you chose to mark a question as understood.