Interview-Day Battle Plan

You have one evening and the interview is tomorrow. This page is the orchestration. Work it top to bottom. It assumes nothing — even if you cram nothing else, read this and the Concepts Rapid-Fire.

Table of Contents

1. The tonight schedule (4–5 hours)

TimeDoWhy
30 minThis page, twiceStructure + the lines you'll actually say
90 min01 — Concepts Rapid-Fire, out loudThese are the screening questions; saying them beats reading
60 min02 — System-Design WalkthroughsInternalize one reference architecture + 4 variants; be able to draw it
45 min04 — Azure Services Deep-DiveThe "do you actually use this?" service questions
30 min03 — Behavioral & BankingSTAR stories + the "is it safe for a bank?" answers
15 min§5 below + track §6 numbers until automaticRecall, not recognition

If you have more time: do Lab 01 hands-on so you can say "I built this last night."

Sleep. A rested brain recalls; a fried one blanks.

2. The first two minutes

They'll open with "tell me about yourself." Have a 60-second answer memorized that hits their MUSTs in their words. Template (fill with your real experience):

"I'm an AI engineer with [X] years building production GenAI on Azure. Most recently I built [agentic / RAG / document-intelligence] systems — for example [one concrete system]: Azure OpenAI for reasoning, Azure AI Search for hybrid retrieval, LangGraph for the agent orchestration with human-in-the-loop, all shipped via FastAPI + Docker + Container Apps with CI/CD and an eval gate. I care a lot about doing this safely in regulated environments — grounding answers, sourcing live data from systems of record rather than the model, and full auditability — which is why [this banking role] is exactly what I want to do next. I also work across Arabic and English, which I know matters here."

Rules: name their stack in their words (Azure OpenAI, AI Search, Document Intelligence, Foundry, LangGraph). Lead with one concrete system, not a list of skills. Signal banking-safety awareness early — it's their #1 filter. Mention Arabic if true.

If your real experience is lighter on banking specifically, bridge: "I haven't worked inside a bank, but I've built to the same bar — grounding, access-trimmed retrieval, human-in-the-loop on high-risk actions, and audit — and I've studied how SR 11-7-style model governance and data residency shape these systems." Honesty + demonstrated understanding beats a bluff every time.

3. The mental model to draw on any whiteboard

Memorize the 8-layer stack. When asked "how would you build X," draw it and walk the layers: Channel → API (FastAPI) → Agent (LangGraph) → Reasoning (Azure OpenAI) → Retrieval (AI Search hybrid) → Ingestion (Document Intelligence/Vision) → Data (Cosmos) → Trust (Entra/Private Endpoints/Content Safety) → Delivery (CI/CD). Drawing it unprompted is the single strongest signal you can send.

The one-liner that frames everything: "The 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."

4. The answer template that wins in banking

For every technical question, append the safeguard. Structure:

  1. Direct answer / design (the how).
  2. The trade-off you considered (shows seniority — there's always a Pareto choice).
  3. The banking safeguard (grounding / tool-not-model / access-trim / human-in-loop / audit / residency).

Example — "How do you build the chat assistant?" → (1) hybrid RAG + LangGraph agent on Azure OpenAI; (2) hybrid+semantic over pure vector because exact identifiers + paraphrase both matter, trading a bit of cost/latency for recall; (3) live balances via a core-banking tool not the model, security-trimmed retrieval, groundedness eval gate, and full audit. That third beat is what they're listening for.

5. Ten facts to never get wrong

  1. Azure Cognitive Search → Azure AI Search; Form Recognizer → Document Intelligence; Azure AI Studio → Azure AI Foundry (now "Microsoft Foundry"). Using old names dates you.
  2. You call a deployment name, not a raw model name, in Azure OpenAI.
  3. Hybrid (BM25+vector, RRF) + semantic reranker beats pure vector on enterprise data.
  4. Never let the LLM be the source of truth for a money fact — tool-call the system of record.
  5. Security-trim retrieval with filters — the LLM is never your access control.
  6. PAYG vs PTU: PTU = reserved, predictable latency, for scale; PAYG = shared, bursty, for dev.
  7. Managed Identity + Key Vault + Private Endpoints — no keys in code, no public egress.
  8. Cosmos: partition key is make-or-break; throughput is RU/s; default consistency Session.
  9. Groundedness/faithfulness is the headline RAG metric; gate releases on it; monitor online.
  10. Microsoft Agent Framework = Semantic Kernel + AutoGen converged (2025), Foundry-native; SK/AutoGen are now predecessors. AI-102 retires 2026-06-30.

6. Land-mines and how to step around them

  • Saying "ChromaDB/Pinecone" for the vector store. This is an Azure role → say Azure AI Search (mention Cosmos vector for colocated memory). Naming a non-Azure store as your default reads as "hasn't done this on Azure."
  • Proposing to fine-tune for the bank's policies/facts. Wrong tool → RAG/tools; fine-tune only for stable style/format.
  • Letting the agent move money autonomously. Always human-in-the-loop on high-risk/irreversible actions.
  • "The model won't reveal other customers' data because I told it not to." No — access control in retrieval/tools/RBAC.
  • Over-engineering multi-agent. Start with one well-scoped agent; add specialists only when justified.
  • Ignoring cost. Mention model tiering, caching, context trimming, PTUs — banks scale to millions.
  • Forgetting Arabic. It changes tokenization, OCR, dialects, RTL, and eval — say so.
  • Hand-waving safety. Be specific: Content Safety, Prompt Shields, PII redaction, audit, residency.
  • Claiming experience you don't have. Bridge honestly (see §2). Interviewers smell bluffing; they reward demonstrated understanding.

7. Questions to ask them

Asking sharp questions signals seniority and that you'll do the job well:

  • "Is the assistant customer-facing or internal/employee first? That changes the grounding and safety bar."
  • "What's the regulatory and data-residency posture — region, on-prem constraints, abuse-monitoring opt-out?"
  • "Are you on Azure AI Foundry Agent Service for orchestration, or self-hosting LangGraph? Any view on Microsoft Agent Framework?"
  • "What does model governance / validation look like here — is there an SR 11-7-style process?"
  • "What's the Arabic scope — dialect handling, Arabic-first or bilingual, and how do you evaluate it?"
  • "What's the biggest production pain today — retrieval quality, hallucination, cost, latency, or eval?"
  • "How do you measure success — groundedness, deflection rate, STP rate, cost/conversation?"

8. Morning-of checklist

  • Re-read §5 and your 60-second intro.
  • Have two concrete stories ready (one win, one failure/lesson) in STAR form (03).
  • Be able to draw the stack (§3) from memory.
  • Water, quiet space, camera/screen-share tested if remote. A whiteboard or paper handy.
  • One line to open with confidence: "I build Azure-native, agentic, grounded GenAI for regulated environments — happy to go as deep as you want on any of it."

9. The rest of this folder