Cheat-Sheet — One Page for Interview Day

Print this. It's the track numbers + ten facts + service map, condensed. Deeper drills: interview-prep/.

The reference architecture (draw this)

Channels → APIM (auth/throttle/audit) → FastAPI (async, SSE) → Agent (LangGraph/
Foundry) → Azure OpenAI (reason) + Azure AI Search (hybrid retrieve) + Tools
(core-banking) → Cosmos (state) ; Doc Intelligence/Vision/OpenCV (ingest) ;
Entra+Managed Identity+Private Endpoints+Content Safety (trust) ; CI/CD+eval gate
+Bicep+ACR+Container Apps+App Insights (deliver)

One-liner: "LLM reasons and phrases; facts come from retrieval and tools; decisions and money stay governed by rules and humans; everything is grounded, evaluated, and audited — because it's a bank."

Current service names (don't use the old ones)

Ten facts to never get wrong

1. You call a deployment name, not a model name, in Azure OpenAI.
2. Hybrid (BM25+vector, RRF) + semantic reranker beats pure vector on enterprise data.
3. Never let the LLM be the source of truth for a money fact — tool-call the system of record.
4. Security-trim retrieval with filters — the LLM is never your access control.
5. PAYG = shared/bursty (dev); PTU = reserved/predictable latency (prod at scale).
6. Managed Identity + Key Vault + Private Endpoints — no keys in code, no public egress.
7. Cosmos: partition key is make-or-break; throughput is RU/s; default consistency Session.
8. Groundedness/faithfulness is the headline RAG metric — gate releases on it, monitor online.
9. LangGraph > a bare loop: durable state (checkpointer) + human-in-the-loop (interrupts) + audit.
10. AI-102 retires 2026-06-30; Microsoft Agent Framework = SK + AutoGen converged, Foundry-native.

Numbers

• token ≈ 4 chars EN / ~2× for Arabic · output tokens cost ~4× input · text-embedding-3-large = 3072-dim.
• 429 = TPM/RPM exceeded → backoff + jitter honoring Retry-After, then PTU/multi-region.
• cost levers: model tiering (mini routes) → prompt caching → trim context → PTU reservations. (sizing tool)

RAG vs fine-tuning vs tools

• RAG/tools → knowledge that changes or is private (policies, balances) — with citations/audit.
• Fine-tuning → stable style/format only — never changing facts.
• Tool call → live/exact data + actions (balance, transfer) — model phrases, never invents.

The 5 banking safety answers (lead with the safeguard)

1. Never a wrong balance? → tool call to core banking; model only phrases; groundedness eval; audit.
2. No cross-customer leakage? → security-trimmed retrieval + OBO auth + RBAC; not the prompt.
3. Prove it to a regulator? → immutable end-to-end audit (prompt/retrieval/tools/decision/approver).
4. Jailbroken / injected doc? → Prompt Shields + least-privilege tools + human approval (architecture limits blast radius).
5. No investment/legal advice? → scoped system prompt + intent routing + refusal-correctness eval.

The labs (all runnable offline)

01 hybrid RAG · 02 LangGraph agent · 03 Document Intelligence · 04 OpenCV/Arabic · 05 FastAPI+Cosmos+CI/CD · 06 eval+guardrails · 07 capstone · 08 NL→SQL · 09 Streamlit UI · 10 MCP+Semantic Kernel. → labs/

60-second intro skeleton

"AI engineer, [X] years production GenAI on Azure. Built [agentic/RAG/IDP] — e.g. [one system]: Azure OpenAI + AI Search hybrid retrieval + LangGraph with human-in-the-loop, on FastAPI + Docker + Container Apps with CI/CD + eval gate. I care about doing this safely in regulated environments — grounding, tools over model memory, auditability — which is why [this banking role] fits. I also work Arabic + English."