Genpact & EXL: GenAI Data Scientist

Answer (genuine and specific):

• "I've seen GenAI move from demos to measurable impact. In my project, a RAG assistant cut document-search time by 60%. I enjoy the mix of NLP, engineering and product thinking it needs."
• "At EXL, the analytics and operations work for insurance, banking and healthcare clients is where GenAI can automate document-heavy processes such as claims, underwriting and KYC, with clear ROI."
• Show you understand the risks (hallucination, privacy) and how you'd manage them.

• The steps: 1. Text is tokenised (BPE / SentencePiece) into IDs, then embedded. 2. Decoder blocks with causal self-attention produce logits over the vocabulary for the next token. 3. Softmax turns the logits into probabilities. 4. Decoding picks a token: greedy, beam search, or sampling with temperature, top-k and top-p. 5. The new token is appended and the process repeats until a stop token or the maximum length.
• A KV cache stores past keys and values so each new token costs O(n) attention rather than recomputing everything.

• OCR + layout parsing for scanned forms (Document Intelligence, Textract).
• Named-entity recognition for policy number, dates, amounts, diagnosis codes: spaCy or a fine-tuned BERT, or an LLM with a JSON schema (zero- or few-shot).
• Classification of claim type and fraud flags; summarisation of adjuster notes.
• Validation: regex and checksum rules, cross-checks against the policy database, and human review when confidence is low.
• Metrics: field-level precision and recall, straight-through-processing rate, handling-time saved.

• Dropout: during training, randomly zero a fraction p of activations, so the network can't rely on specific neurons. It acts as an ensemble of sub-networks and reduces overfitting. At inference it is off (activations are scaled accordingly).
• Batch Norm: normalises a layer's inputs per mini-batch (mean 0, variance 1), then applies a learned scale and shift. It stabilises and speeds up training and allows higher learning rates.
• Transformers use LayerNorm instead (per token, independent of batch size), usually as pre-norm in modern LLMs.

Answer (a structure that works):

1. Context and goal (1 line, with the metric).
2. Data: sources, size, quality issues and the fixes.
3. Approach: baseline → final model, and why; the alternatives you rejected.
4. Evaluation: offline metrics plus a business check.
5. Deployment: how it runs, how it's monitored.
6. Impact and learnings: numbers, plus what you'd do differently. - Keep it to about 3 minutes, then let the interviewer dig in.

def dedupe(xs):
    seen = set()
    return [x for x in xs if not (x in seen or seen.add(x))]

print(dedupe([3, 1, 3, 2, 1]))   # [3, 1, 2]
# or: list(dict.fromkeys(xs))   # dicts keep insertion order (Python 3.7+)