Company Interview Questions · Banking & Fintech

Citi: Data Scientist & GenAI

❓ 6 questions🏢 Citi🆕 Recent reports

Citi: Data Scientist & GenAI Interview Experiences (2024–26). Rounds, real reported questions and model answers, compiled from public candidate reports (sources at the end). Practice mode hides every answer: say your answer first, press Show answer, then rate yourself. Use 🔁 Review again to build your revision list and 🎲 Random question for a quick drill.

★ Must-know 145 📝 Assignments

About this interview

Compiled from four public candidate reports for Citi Data Scientist roles: Bengaluru (November 2025, GenAI/RAG focus), December 2025 (offer accepted), Shanghai (April 2026, case-study presentation) and October 2024. Citi's India GenAI openings (e.g. Senior Data Scientist – Gen AI Engineer, Pune and Bengaluru) give context. The questions are paraphrased, and the model answers are written by EasyLearnSpace.

Roles	Data Scientist; Senior Data Scientist – Gen AI Engineer (AVP); Gen AI Analyst
Rounds	Online aptitude / reasoning / data-interpretation test → 2 technical rounds → HR. Other locations: technical → case-study presentation → values and leadership fit
Reported topics	Decision trees, RAG system design, vector DBs, LLMs, batch normalisation, optimising an RL algorithm, ML and DL coursework, resume projects
Outcomes	1 offer (rated easy, "terminology heavy"); 3 not selected (Bengaluru rated difficult)

01Rounds at a glance

Technical rounds combine classic ML (trees), deep learning (batch norm) and GenAI system design (RAG + vector DB).
Banking context matters: explainability, data privacy and model risk governance.

02Questions and model answers

How does a decision tree decide where to split? How do you stop it overfitting?

Classification: choose the feature and threshold that maximise impurity reduction.
Gini = 1 − Σpᵢ², or entropy = −Σpᵢ log pᵢ, weighted by child sizes (information gain).
Regression: minimise the variance (MSE) of the children.
Greedy and recursive: it continues until a stopping criterion is met.
Control overfitting:
max_depth, min_samples_leaf, min_samples_split.
Cost-complexity pruning (ccp_alpha).
Or ensembles: random forests (bagging) and gradient boosting.
In banking: shallow trees and scorecards are favoured for explainability to regulators.

Design a RAG system for internal banking policies and procedures.

Ingestion: policy PDFs and SOPs → layout-aware parsing → clause-level chunks with metadata (policy ID, version, effective date, business unit, classification) → embeddings → vector DB, plus BM25.
Access control: entitlement filters at retrieval (users see only the policies for their business unit and classification).
Query: rewrite → hybrid search → cross-encoder re-rank → latest effective version only → grounded answer with clause citations.
Guardrails:
No financial advice to customers.
PII redaction.
"Not found" when retrieval confidence is low.
Prompt-injection filtering.
Governance:
A model inventory entry and validation (model risk management).
An evaluation set signed off by compliance.
Audit logs, periodic review, and data residency on a private model endpoint.

What is a vector database and how would you choose one for a bank?

What it does: stores embeddings with ANN indexes (HNSW, IVF-PQ) for fast similarity search, plus metadata filtering.
Bank criteria:
On-premises or private-cloud deployment.
Encryption at rest and in transit.
RBAC and attribute filters.
Audit logs.
High availability and backups.
Scale (millions of chunks).
Hybrid search.
Vendor approval.
Options: pgvector (inside an approved Postgres), Elasticsearch/OpenSearch (often already approved), Milvus, Qdrant (self-hosted), Azure AI Search with private endpoints.

Explain batch normalisation and why it helps.

For each mini-batch, normalise the activations to mean 0 and variance 1, then apply a learned scale γ and shift β.
Benefits:
Stabilises the distribution of layer inputs.
Allows higher learning rates and faster convergence.
Reduces sensitivity to initialisation.
Has a mild regularising effect.
Inference: uses running averages of the mean and variance.
Limitations: depends on batch size (small batches hurt), so Transformers use LayerNorm.

How would you optimise a reinforcement-learning algorithm that trains slowly or unstably?

Stability:
Target networks and experience replay (DQN).
Clipped objectives (PPO).
Gradient clipping.
Normalised observations and rewards.
Entropy bonus for exploration.
Sample efficiency:
Prioritised replay.
Off-policy algorithms (SAC, TD3).
Reward shaping (carefully).
Imitation learning or pre-training on logged data.
Speed: vectorised and parallel environments, a smaller network, tuned hyperparameters (learning rate, γ, batch size), and tracking learning curves over several seeds.
Banking use cases: dynamic pricing, collections strategy, trade execution. Mention offline RL and the need for safe policy evaluation.

Case-study presentation: how do you structure a 15-minute data-science case presentation for senior stakeholders?

Executive summary first (1 slide): the problem, the recommendation, the impact.
Context and data (2 slides).
Approach (2–3 slides): keep it simple, explain why.
Results (2 slides): metrics and business impact in money terms.
Risks and next steps (1 slide): limitations, monitoring, governance. - Leave time for questions. Executives probe assumptions and risk, not hyperparameters.

03How to prepare for Citi

Revise classic ML (trees, ensembles) and DL basics (batch norm, optimisers).
Prepare a bank-grade RAG design with entitlements, citations and model risk governance.
Expect an aptitude and data-interpretation online test in India.

Sources

Created by Ajeetkumar