MangaAssist Interview Pack - Medium With Hints

Level: Medium
How to use: Use the hints to pressure-test whether your answer includes routing logic, data choices, and the documented tradeoffs.

Flow Map

graph LR
    A[User Message] --> B[Intent Classifier]
    B -->|chitchat| C[Template Path]
    B -->|order tracking| D[Order API]
    B -->|product question| E[Catalog Lookup]
    B -->|recommendation| F[Reco Engine + Catalog]
    B -->|faq / policy| G[RAG Pipeline]
    B -->|ambiguous| H[Full LLM Path]
    C --> I[Guardrails]
    D --> I
    E --> I
    F --> I
    G --> I
    H --> I
    I --> J[Response]

Interview Questions With Hints

Product Manager

1. If a user says, “I liked Demon Slayer, what should I read next?”, what signals from this project would you use before generating the response?

Hint: Use the seed title, possible user history, current page context, recommendation signals, and maybe editorial/RAG enrichment.

1. How would you explain the difference between product discovery, recommendation, and product Q&A in user-facing terms?

Hint: Discovery is broad exploration, recommendation is similar-item guidance, and Q&A is precise fact lookup.

Senior Engineer

1. Describe the full flow of a recommendation request from the moment the user sends it to the moment the UI displays the answer.

Hint: Include memory load, intent classification, parallel service fan-out, prompt building, LLM generation, and guardrails.

1. Why does MangaAssist use a two-stage intent classification path instead of a single model for every message?

Hint: Rules are faster and cheaper for obvious patterns; the model handles ambiguity; full-LLM routing for everything would be slower and costlier.

1. What is stored in conversation memory, and why is TTL important here?

Hint: Session turns, page context, last intent, maybe user context. TTL supports privacy and cost control.

1. Why are prices treated differently from product details, promotions, and reviews in the caching strategy?

Hint: Prices are freshness-critical and should be fetched live to avoid trust-damaging stale data.

ML Engineer

1. What is the point of reranking in the RAG pipeline, and why is vector similarity alone not enough?

Hint: Top-k vector retrieval gets candidates; cross-encoder reranking improves semantic precision before prompt injection.

1. How does the design try to reduce hallucinations when the LLM talks about products, prices, or policies?

Hint: Grounding, structured product data, explicit prompt rules, lower temperature, ASIN/price validation.

SRE

1. What parts of this architecture are critical versus nice-to-have during a partial outage?

Hint: Tier 1 keeps the chatbot alive, Tier 2 reduces capability, Tier 3 is optional.

Analytics Lead

1. How would you connect operational metrics like latency and guardrail block rate to business metrics like conversion and support deflection?

Hint: Explain causal chains, not just metric lists. Poor latency and high blocks reduce resolution and purchase confidence.

Sequence Recall

sequenceDiagram
    participant User
    participant Orch as Orchestrator
    participant Memory as DynamoDB Memory
    participant Intent as Intent Classifier
    participant Reco as Recommendation Engine
    participant Catalog as Product Catalog
    participant LLM as Bedrock LLM
    participant Guard as Guardrails

    User->>Orch: Recommend manga like Vinland Saga
    Orch->>Memory: Load session context
    Orch->>Intent: Classify intent
    Intent-->>Orch: recommendation
    par Data fan-out
        Orch->>Reco: Get ranked titles
        Orch->>Catalog: Get product metadata
    end
    Reco-->>Orch: Ranked ASINs
    Catalog-->>Orch: Product details
    Orch->>LLM: Prompt with structured context
    LLM-->>Orch: Draft answer
    Orch->>Guard: Validate output
    Guard-->>User: Final response