Citi bank

Round 0 Type: Written Round This round consisted of Logical Reasoning questions, Case Study–based questions, and many questions from Statistics and Probability. If you know Mathematics-3 well and remember most of its formulas, you can perform very well in this round. -- Round 1 Type: Technical Interview The interview was online. The interviewer introduced himself as the VP of the Engineering Team at Citi. He went through my resume and mentioned that my profile looked very strong in Machine Learning and more oriented towards academic research. He asked why I wanted to work at Citi and whether I planned to stay there long term. I explained about my current research internship and its drawbacks, and also mentioned funding issues for Indian students, which is why I wanted to move into the corporate sector in India. He then noticed my summer internship at IBM Research and asked about the project, my responsibilities, and overall working experience. He also asked about Retrieval Augmented Generation (RAG). Next, he gave me an open-ended question: how would I use Generative AI to solve current problems at Citi and improve efficiency of its products. I answered by proposing ideas such as using agents to monitor stock prices, detect data drift, and notify ML engineers, along with feedback mechanisms similar to RLHF to improve model performance. He seemed impressed by the answer. The interview lasted around 15–20 minutes. Before ending, he asked one final “Why Citi?” question and concluded the round. -- Round 2 Type: Technical Interview I was the first candidate to attend this round. The interviewer explained that the interview would proceed in three parts: ML basics, resume discussion, and coding. He began with Machine Learning fundamentals: Asked about multicollinearity Questions on Linear Regression and Logistic Regression Underlying assumptions of Linear Regression Decision Trees and Random Forests Given a scenario with 100 elements vs 500 elements in leaf nodes, asked which model is simpler Hyperparameters in ML models (I explained number of trees in Random Forest) Learning rate and gradient descent behavior He asked: if learning rate is 0.5 vs 0.1, which converges faster and which is better. I explained that a larger learning rate may converge faster but risks missing the global minimum, while a smaller one converges slower but more safely. Then he moved to evaluation metrics: Precision, Recall, F1 Score (with mathematical formulas) Next scenario: evaluating a French-to-English Seq2Seq model. I suggested BLEU score and explained how it works. He asked when BLEU can give misleadingly high scores. I explained cases like hallucinated or over-detailed outputs. He asked about embeddings: I suggested Word2Vec, but he wanted simpler ones. He hinted toward TF and IDF; I explained the concepts but mentioned I hadn’t used them practically. He then asked about ROC and AUC; I admitted I knew the concepts but wasn’t confident with the graph interpretation. After that, he asked me to explain the evolution of Neural Networks from Perceptron to Transformers and the advantages and challenges at each stage. Coding Section: He asked SQL queries verbally: One Self Join problem One GROUP BY and HAVING problem Find the second maximum salary I solved them verbally. For the last question I didn’t use RANK; he asked if I could, but I didn’t know the syntax and he was fine with it. Then he asked a simple Python problem to divide a dataset into three ranges using lambda if possible. I began writing using a loop, but he stopped and said the interview was complete since I had answered most of what he expected. After this round, I came to know I was selected. Many candidates had an HR round, but mine was mostly a formality. -- Round 3 Type: HR Round The HR asked why I was in Singapore for my internship and whether I would continue working there after it. She asked about my family background, career goals, and location preference. They had an office in Hebbal (Bangalore). I said I was comfortable with both Chennai and Bangalore, and she mentioned internal transfer to Chennai after a year was possible. I was already selected before this round; it was just formal documentation and confirmation. -- Experience The first interviewer was relaxed, making it my shortest interview. The second round lasted nearly 80 minutes and covered a wide range of Machine Learning topics along with SQL. Even though I received hints when stuck and didn’t feel very confident afterward, I was unexpectedly selected. For future ML interviews, reviewing fundamental ML questions is helpful. For SQL, practicing questions related to joins and aggregate functions is recommended.