Tech Mahindra: AI Engineer, Noida (Offer)

• Business problem → data → model → serving → monitoring.
• Example: "An email-triage assistant for a telecom client.
• Emails go through PII masking and are then classified into 12 intents by a fine-tuned DistilBERT.
• For complex cases, a GPT-4o-mini RAG step drafts a reply from the knowledge base.
• The service runs on FastAPI in Azure App Service and is monitored with Application Insights.
• Outcome: 38% fewer manual touches and 2 minutes saved per email."
• Always say why each component was chosen and what you measured.

• Embeddings: tokens become embeddings, and positional encodings are added.
• N blocks, each containing multi-head self-attention (softmax(QKᵀ/√d)V) and a feed-forward network, with residual connections and layer norm around both.
• Variants: encoders (BERT) understand text, decoders (GPT) generate it with a causal mask, and encoder-decoders (T5) translate or summarise.
• Why it is powerful: parallel training, long-range context, and it scales to LLMs.

• Models: Azure OpenAI (private endpoint) or Azure AI Foundry for open models.
• Retrieval: Azure AI Search (hybrid + semantic ranker), with Blob Storage for documents.
• Compute: Azure Container Apps, AKS or App Service running FastAPI; Functions for ingestion jobs.
• Security: Entra ID authentication, Key Vault for secrets, VNet with private endpoints, and Content Safety filters.
• Operations: Application Insights / Azure Monitor, plus a CI/CD pipeline in Azure DevOps or GitHub Actions.
• On AWS: Bedrock, OpenSearch, ECS or Lambda, Secrets Manager, CloudWatch.

def primes_upto(n):
    if n < 2: return []
    sieve = [True] * (n + 1); sieve[0] = sieve[1] = False
    for i in range(2, int(n ** 0.5) + 1):
        if sieve[i]:
            sieve[i*i::i] = [False] * len(sieve[i*i::i])
    return [i for i, p in enumerate(sieve) if p]

• The Sieve of Eratosthenes runs in O(n log log n). Return an empty list when n < 2.

• Label: the customer churned within the next 30 or 60 days.
• Features: tenure, plan, ARPU trend, data usage drop, complaints and support calls, payment delays, network quality in the customer's area.
• Model: a gradient-boosted tree with time-based validation, judged on PR-AUC, and lift in the top decile.
• Action: give retention offers to the high-risk, high-value segment; measure uplift with a control group, not just accuracy.
• Explain the drivers with SHAP for the business team.

import pandas as pd
df = pd.read_csv("customers.csv")
print(df.isna().mean().sort_values(ascending=False))            # missing ratio
df["income_missing"] = df["income"].isna().astype(int)          # keep the signal
df["income"] = df["income"].fillna(df["income"].median())
df["city"] = df["city"].fillna("Unknown")
df = df.dropna(subset=["customer_id"])                          # critical key

• Strategy: understand why values are missing (completely at random, at random, or not at random). Drop a column if more than ~60% is missing and it isn't useful. Impute with median or mode, or use model-based imputation (KNN, iterative). Fit imputers on the training data only.