My LinkedIn Connection Just Cleared Genpact's GenAI Lead Interview — Round 2 Full Breakdown With Every Answer
15 questions across GenAI, RAG, ML math, statistics, and live coding. Principal Consultant level. Here’s how she answered every single one.
She messaged me on a Tuesday afternoon with three words: “I think I cleared.”
Priya has been in data science for 8 years — the last three deeply focused on GenAI and agentic systems. She’d applied for a GenAI Lead (Principal Consultant) role at Genpact, cleared Round 1 screening, and had just wrapped up Round 2 — the one that actually matters.
“It wasn’t just GenAI questions,” she told me. “They went deep into ML math, statistics, and even asked me to draw cosine similarity on a virtual whiteboard. I wasn’t expecting that at a Principal Consultant level.”
She shared every question with me and said — “Post this. Someone needs it.”
So here it is. Every question, exactly how Priya answered it.
Career Introduction
This is never just a formality at the Principal Consultant level. The interviewer is assessing your career narrative — does your journey make sense? Do you own a clear progression?
Priya structured hers in three acts:
- Foundation (Years 1–3): Started as a Data Analyst, moved into ML Engineering. Built classical ML models in banking and retail. Learned to productionize models, not just build them.
- Transition (Years 4–6): Shifted into NLP and deep learning. Led NLP projects — entity extraction, document classification, semantic search. First exposure to LLMs via GPT-3 fine-tuning.
- GenAI Leadership (Years 6–8): Built and led a 6-person GenAI team. Delivered 3 production agentic AI systems for enterprise clients. Currently advising Fortune 500 clients on AI strategy.
Tell me about your current project in detail — client requirement, technical architecture, tech stack, and your individual responsibilities.
This was the longest question — Priya spent 12 minutes on it. At the Principal Consultant level, they want to see that you own the solution, not just contribute to it.
A large insurance company was drowning in ServiceNow tickets — 10,000+ incidents per month. L1 agents spent 60% of their time on repetitive RCA tasks. The client wanted an agentic AI system to triage, analyze, and resolve Tier-1 incidents autonomously.
- Ingestion Layer: 5 years of historical incident data + 200+ SOPs embedded using text-embedding-3-small, stored in pgvector.
- Agentic Layer: LangGraph-based multi-agent system with a Supervisor Agent routing to 4 specialist sub-agents: SOP Retriever, Incident Classifier, Resolution Generator, ServiceNow Updater.
- RAG Pipeline: Hybrid retrieval (dense + BM25) → cross-encoder re-ranking → top-3 chunks injected into GPT-4o prompt.
- Feedback Loop: Human agents flagged incorrect resolutions → bi-weekly re-ranker retraining.
- Deployment: FastAPI on GCP Cloud Run, API Gateway with API key auth, Secret Manager.
LangGraph · OpenAI GPT-4o · pgvector · PostgreSQL · FastAPI · GCP Cloud Run · API Gateway · Python · Streamlit
- End-to-end solution architecture — requirements to production deployment.
- Designed multi-agent orchestration logic and state management.
- Built hybrid retrieval pipeline (30% improvement in context precision vs naive RAG).
- Led client demos and stakeholder presentations. Managed 3 developers.
How many total agents were in your solution and how were they orchestrated?
Total agents: 5 (1 Supervisor + 4 Specialist Agents)
| Agent | Responsibility | Triggers When |
|---|---|---|
| Supervisor Agent | Reads state, decides next agent | Entry point and after every completion |
| SOP Retriever | Fetches relevant SOPs from pgvector | sop_context is empty in state |
| Incident Classifier | Classifies severity: P1/P2/P3 | Severity not yet set in state |
| Resolution Generator | Generates resolution steps via RAG | SOP retrieved + severity classified |
| ServiceNow Updater | Pushes resolution to ticket | Resolution generated + human approved |
Orchestration: LangGraph StateGraph with conditional edges on the Supervisor node. Human-in-the-loop approval gate before the ServiceNow Updater.
How did you make sure AI responses were from internal resources only and not based on publicly available information?
Priya described a 3-layer containment strategy:
No query reaches the LLM without retrieved internal context attached. If retrieval returns zero relevant chunks (similarity score below threshold), the query is blocked at the retrieval layer — the LLM is never called.
After generation, a faithfulness scorer (RAGAs) checks every claim. Responses below 0.85 faithfulness are blocked and replaced with a safe fallback message.
How did you evaluate your RAG solution?
Priya used the RAGAs framework — and the interviewer asked her to name every metric and explain what each measures.
| Metric | What It Measures | Target |
|---|---|---|
| Context Precision | Of retrieved chunks, what % were actually relevant? | > 0.80 |
| Context Recall | Of all relevant info, what % did we retrieve? | > 0.75 |
| Faithfulness | Is every claim supported by retrieved context? | > 0.85 |
| Answer Relevance | Does the answer address the question? | > 0.80 |
| Answer Correctness | Match against gold standard answers | > 0.75 |
Beyond RAGAs: 200 golden QA pairs built by domain experts, LLM-as-Judge scoring via GPT-4o, and operational metrics (latency, token cost, cache hit rate) tracked in Looker dashboards weekly.
What if your system is not able to find relevant data in the corpus? How will it respond?
Priya described a tiered fallback strategy:
- Similarity threshold check: If top chunk cosine similarity < 0.65, the LLM is never called. Response: “I couldn’t find relevant information in our knowledge base.”
- Partial match fallback: Similarity 0.65–0.75 → respond with a confidence disclaimer: “Based on the closest matching info I found — please verify with your team lead.”
- Escalation routing: Unanswered queries are automatically routed to an L2 queue in ServiceNow.
- Corpus gap logging: Every unanswered query logged in a “knowledge gap” table — feeds a monthly SOP review cycle.
How will you handle it if your RAG solution is giving wrong answers confidently?
- Faithfulness scoring (RAGAs): Automated check after every generation — unfaithful answers are caught and blocked.
- Dual-LLM cross-check: A second LLM independently asked: “Does this answer contradict the provided context?” Disagreements trigger a block.
- Human feedback loop: Reviewers flag wrong resolutions → stored for re-ranker retraining.
- Citation enforcement: Every claim must reference a chunk ID like [SOP-047, Section 3.2]. A validator checks all citations exist in retrieved chunks.
- Temperature = 0: Set for factual generation — reduces creative/hallucinatory output.
- Chunking audit: Confident wrong answers often trace to bad chunks. Retrieval logs audited weekly.
What is cosine similarity? What is the mathematical background? (Virtual whiteboard)
Priya said this was the most unexpected question. She was glad she understood it geometrically, not just as a formula.
Cosine similarity measures the angle between two vectors — not the distance between their endpoints. Two vectors pointing in the same direction score 1 (identical), perpendicular vectors score 0 (unrelated), opposite vectors score −1.
Why angle instead of distance? In NLP we care about meaning (direction), not magnitude. A 3-word sentence and a 300-word document on the same topic should be considered similar even though their vector magnitudes differ wildly.
Why it matters for RAG: When a user query is embedded into a 1536-dimensional vector, cosine similarity finds chunks whose embedding vectors point in the most similar direction — regardless of text length.
What is p-value and when should we use it?
Definition: The p-value is the probability of observing results at least as extreme as the ones you got, assuming the null hypothesis is true. It does NOT tell you the probability that your hypothesis is correct.
- A/B testing: Is the new model version performing significantly better than the baseline?
- Feature selection: Is this feature statistically significantly correlated with the target?
- Model comparison: Is the performance difference between two models real or just noise?
Which algorithms have you used for classification so far?
Priya grouped them by family and mentioned where she’d actually used each one — not just a list of names.
| Family | Algorithms | Used For |
|---|---|---|
| Linear | Logistic Regression, SVM (SVC) | Churn prediction, text classification |
| Tree-Based | Decision Tree, Random Forest, XGBoost, LightGBM | Fraud detection, incident severity |
| Probabilistic | Naive Bayes | Spam detection, document categorization |
| Instance-Based | KNN | Recommendation, anomaly detection |
| Neural | MLP, Fine-tuned BERT | Sentiment analysis, intent classification |
| Ensemble | Voting, Stacking | High-stakes production models |
Explain the mathematical background of any one ML algorithm. (Priya chose SVC — with 3 follow-up questions)
SVC finds the optimal hyperplane that separates two classes with the maximum margin. The margin is the distance between the hyperplane and the nearest data points (support vectors) from each class.
When data isn’t linearly separable, the kernel trick implicitly maps it into a higher-dimensional space where it becomes separable — without computing those high-dimensional coordinates explicitly. Common kernels: RBF (projects to infinite dimensions), Polynomial, Sigmoid.
- High C: Classifies all training points correctly → small margin → risk of overfitting.
- Low C: Accepts some misclassifications → larger margin → better generalization.
- SVC wins when classes are clearly separable and dataset is small-to-medium.
- LR is faster, more interpretable, scales better to large datasets.
- SVC with RBF beats LR on non-linear boundaries but requires feature normalization.
What’s the difference between covariance and correlation?
| Dimension | Covariance | Correlation |
|---|---|---|
| Measures | Direction of linear relationship | Direction + strength |
| Range | −∞ to +∞ (unbounded) | −1 to +1 (normalized) |
| Units | Product of units of X and Y | Dimensionless |
| Interpretability | Hard to compare across datasets | Easy to compare — standardized |
| Formula | Σ(xᵢ−x̄)(yᵢ−ȳ) / n | Cov(X,Y) / (σₓ × σᵜ) |
What is a confusion matrix, its significance, and what metrics can be derived from it?
| Predicted: Positive | Predicted: Negative | |
|---|---|---|
| Actual: Positive | TP (True Positive) | FN (False Negative) |
| Actual: Negative | FP (False Positive) | TN (True Negative) |
What is collinearity?
Collinearity (multicollinearity) occurs when two or more predictor variables in a regression model are highly correlated with each other.
- Makes it difficult to determine each feature’s individual effect on the target.
- Coefficients become unstable — small data changes cause large coefficient swings.
- Standard errors inflate → hypothesis tests become unreliable.
- Correlation matrix: Pairwise correlations above 0.8–0.9 are a red flag.
- VIF (Variance Inflation Factor): VIF > 5 = moderate; VIF > 10 = severe.
- Remove one of the correlated features.
- Use PCA to create uncorrelated components.
- Apply Ridge Regression (L2 regularization) — shrinks correlated coefficients.
Write a program to subtract two numbers with fractions — assuming your compiler doesn’t understand fractions.
Beautifully deceptive. It tests whether you understand fraction arithmetic at a mathematical level and can implement it from scratch.
The key insight: represent fractions as two integers (numerator + denominator) and implement the math yourself.
What Genpact’s PC Round Is Really About
When I asked Priya what surprised her most, she said: “I expected it to be 90% GenAI. It was 50% GenAI, 30% ML math, and 20% statistics. They really want Principal Consultants who haven’t forgotten their foundations.”
At the PC level, Genpact isn’t just hiring GenAI coders — they’re hiring people who can walk into a client boardroom and defend every architectural and mathematical decision.
- Own your architecture end-to-end. Know why you made every design decision — agents, retrieval, evaluation, fallbacks.
- Don’t neglect ML fundamentals. SVC, confusion matrix, p-value, collinearity — table stakes at PC level.
- The coding question will have a trick. “Fractions without fractions” tests mathematical maturity, not syntax knowledge.
- Graceful degradation > happy path. Every RAG question had a failure-mode follow-up. Show you’ve thought about what breaks.
If this helped your prep — give it 50 claps. It helps more engineers find this content.
Have you interviewed at Genpact, Accenture, or any top firm recently? DM me on LinkedIn — I’ll feature your story next.
Follow me for a new company interview breakdown every single day.