AI Engineering
Learn LLMs, RAG, agents, secure tools, and production AI through practical insurance examples.
Design 2: Customer-support agent
A bounded assistant for knowledge questions, account lookups, clarification, and escalation. It may perform approved low-risk actions.
Design 3: Insurance claim agent
A claim-preparation assistant inside an authoritative claim workflow. It gathers and drafts information without independently adjudicating claims.
Design 4: AI CRM agent
A staff assistant that summarizes interactions and proposes CRM updates or follow-ups. Its tools obey staff role permissions.
Design 5: Document processing system
An asynchronous pipeline turns uploads into validated extracted data and search indexes. It is primarily a workflow with AI stages.
Design 6: Email agent
A mailbox-aware assistant for triage, search, summaries, and reviewed outgoing messages. Sending is a controlled tool.
Design 7: Research agent
A bounded system gathers external evidence and produces a cited research draft. Its output is decision support, not verified universal truth.
Design 8: Multi-agent system
Several specialized model-driven components cooperate under a controlled coordinator. Each has a measurable task and narrow context.
Design 9: Enterprise knowledge assistant
A permission-aware assistant over many organizational sources and users. It requires identity-linked document access, freshness, and trustworthy source references.
The overall insurance assistant
An insurance assistant helps authenticated customers understand documents, check claim status, prepare submissions, and reach staff. It combines RAG, exact service lookups, and bounded agent behavior. It does not become the authoritative policy or…
Which parts are deterministic and which are agentic?
Agentic portion: interpreting wording, gathering missing information, and choosing among bounded read/search actions. Workflow portion: identity, required fields, approvals, transaction order, retries, and durable outcomes. LangGraph can coordinate the…
Tools and their safe contracts
Never accept a model's customerId as proof of the current customer. Never let the model supply an unrestricted email URL, SQL query, database filter, or role field.
Data model
Enforce unique operation keys and optimistic draft versions. Keep large original files in protected object storage. Keep model-visible projections small. A document index and conversation cache are rebuildable derived data; claim and approval records need…
API design
Apply CSRF protection where cookie-authenticated mutation routes require it, enforce origin/authentication policy, and use appropriate error semantics. A pending task may return 202; a chat sentence claiming success cannot substitute for a claim receipt.
Two complete request walkthroughs
A. “Check claim C42.”
Failure handling, evaluation, and operational review
A production reference design specifies failed and uncertain outcomes as carefully as successful ones. Evaluation checks behavior before release; observability detects failures in actual runs. Human escalation is an explicit supported outcome.
Project 1 — Basic LLM chatbot
Build in this order. Each project extends a capability, not necessarily a separate repository. These are future learning projects; the notes remain in this one file. Move forward when you can explain the trade-offs and demonstrate the acceptance criteria,…
Project 2 — Streaming AI chatbot
Requirements: Show progress/text incrementally, allow cancellation, and distinguish completion from stream failure. Preserve final message state.
Project 3 — PDF RAG assistant
Requirements: Upload approved PDFs, track ingestion, ask questions, return source references, and abstain when evidence is missing.
Difficulty is an editorial estimate. Completion means you chose to mark a question as understood.