BCG X: AI/ML Engineer & Forward-Deployed DS

def flatten(obj, parent="", sep="."):
    out = {}
    if isinstance(obj, dict):
        for k, v in obj.items():
            out.update(flatten(v, f"{parent}{sep}{k}" if parent else k, sep))
    elif isinstance(obj, list):
        for i, v in enumerate(obj):
            out.update(flatten(v, f"{parent}{sep}{i}" if parent else str(i), sep))
    else:
        out[parent] = obj
    return out

print(flatten({"a": 1, "b": {"c": 2, "d": [3, 4]}}))  # {'a':1,'b.c':2,'b.d.0':3,'b.d.1':4}

import pandas as pd

total, count = 0.0, 0
for chunk in pd.read_csv("big.csv", usecols=["amount"], chunksize=1_000_000):
    s = chunk["amount"].dropna()
    total += s.sum(); count += len(s)
print(total / count)

• Stream in chunks, keeping a running sum and count. Alternatives: DuckDB (SELECT avg(amount) FROM 'big.csv'), Polars lazy scan, or Spark for distributed data. For a running variance, use Welford's algorithm.

• Ingest: connectors (SharePoint, S3) → file-type detection → OCR for scans (Azure Document Intelligence or Textract) → layout-aware parsing (tables, clauses).
• Enrich: extract metadata (parties, dates, contract type) with an LLM using structured output; detect and redact PII.
• Index: chunks by clause or section → embeddings + BM25 → a hybrid index with metadata filters and document-level ACLs.
• Serve: search UI + RAG Q&A with citations; a re-ranker; "find all contracts with auto-renewal clauses" via metadata plus semantic search.
• Govern: audit logs, retention policy, an evaluation set built with client SMEs; measure time saved per review.

• Real-time scoring at authorisation: features (velocity, merchant risk, geo-distance, device) from an online feature store, a GBM model, a latency budget under 100 ms.
• Decision tiers: approve / SMS confirmation ("Reply YES if this ₹24,999 purchase was you") / decline.
• SMS flow: send through a gateway, hold the transaction for N minutes; YES approves, NO blocks the card and opens a case; no reply follows the default policy.
• Feedback loop: customer replies and chargebacks become labels; retrain regularly; monitor false-positive rates by segment.
• Risks: SMS phishing (never include links), delivery delays, fairness across customer groups.

1. Frame the value: churn rate, revenue at risk, the cost of retention offers. "A 1-point churn reduction ≈ ₹X crore a year."
2. Hypotheses: price sensitivity, poor onboarding, delivery issues, competitor offers.
3. Data: subscriptions, usage, support tickets, NPS, delivery performance.
4. Approach: a churn-risk + uplift model → segment-specific actions (onboarding fixes, targeted offers, service recovery).
5. Metrics: retained revenue, offer ROI, and a controlled pilot in 2 regions.
6. Adoption: who uses the scores (the CRM team), workflow integration, a dashboard, training, and governance for model refresh. - Keep the technical detail light and tie every step to client value.

• "I enjoy turning ambiguous business problems into shipped AI products. A forward-deployed role puts me close to the client's real constraints and users, where impact is measured, not assumed."
• Show evidence: a time you handled ambiguity, brought stakeholders along, and drove adoption (not just model accuracy).
• Ask good questions back: how teams split between BCG X engineers and consultants, and how success is measured after go-live.