Zensar, Birlasoft & GlobalLogic: AI / GenAI Engineer

• Pearson r = cov(X, Y) / (σ_X σ_Y), ranging from −1 to 1. It measures the strength and direction of a linear relationship.
• Limitations:
• Misses non-linear relationships (use Spearman for monotonic ones, or mutual information).
• Sensitive to outliers.
• Correlation ≠ causation.
• Can be distorted by confounders or by combining groups (Simpson's paradox).
• In ML: high correlation between features means multicollinearity, so drop one or use regularisation or PCA.

• Boxes: client/UI → API gateway / auth → app service (FastAPI) → orchestration (LangChain/LangGraph) → LLM provider; retrieval (vector DB + document store); ingestion pipeline (loader → chunker → embedder, on a schedule); cache (Redis); observability (logs, traces, evaluation dashboard).
• Arrows labelled with data flow (query, context, response, feedback).
• Call out: where PII is masked, where secrets live, scaling points and failure fallbacks.
• Practise drawing it in 3 minutes on a whiteboard or in Excalidraw. Interviewers asked for exactly this.

1. First 15 minutes: clarify the scope and pick a narrow, demo-able feature. Sketch the UI and API.
2. Backend: FastAPI with one or two endpoints and an LLM call with structured output; mock the data if needed.
3. Frontend: Streamlit, or a simple React/HTML form, just enough to demo.
4. Integrate early (by hour 1.5), then improve error handling and the UX.
5. Last 30 minutes: a README, a demo script, and known limitations. Commit to git frequently. - Judges reward a working end-to-end demo over unfinished ambition.

def has_cycle(head):
    slow = fast = head
    while fast and fast.next:
        slow, fast = slow.next, fast.next.next
        if slow is fast:
            return True
    return False

• Floyd's tortoise and hare: O(n) time, O(1) space. A follow-up asks you to find the cycle's start: reset one pointer to the head and move both one step at a time until they meet.

1. Confirm and scope: dashboards (p50/p95), which endpoints, all regions or one.
2. Correlate with the change: diff the release; roll back if the impact is severe (mitigate first).
3. Trace: distributed traces show the slow span (DB query, LLM call, external API).
4. Common causes: an N+1 query, a missing index, cache disabled, a larger prompt (more tokens), synchronous calls that were parallel before, the connection-pool limit.
5. Fix, verify and prevent: a load test in CI, alerts on SLO breach, and a postmortem.

• Model (chat LLM wrapper), prompt template, output parser, retriever, tool, and runnable chains (LCEL) composed with |.
• LangGraph adds stateful graphs for agents: nodes, edges, loops and checkpoints.
• Supporting tools: LangSmith for tracing and evaluation, and integrations for 100+ vector stores and loaders.
• Value: faster prototyping and swappable providers. Caveat: abstraction overhead, so keep critical paths simple.