The most comprehensive bank Interview Answer Review tooling available online.

select sum(amount)/count( distinct transaction_id) as avg_transaction

FROM transactions WHERE transaction_date >= CURRENT_DATE - INTERVAL '90 days'

GROUP BY customer_id

HAVING COUNT(transaction_id) > 10;

At my previous role, I was leading a fraud detection AI model deployment for a financial services client. The model aimed to reduce fraudulent transactions by 25% while maintaining a low false positive rate.

I was confident in the model’s performance in the test environment, but when we deployed it in production, we noticed an unexpected spike in false positives. Transactions flagged as fraudulent were legitimate, causing disruptions for customers and increasing manual review work. To resolve this, I:

1. Investigated the issue by comparing training and production data.
2. Identified data drift—the model was trained on historical data, but recent transaction patterns had changed.
3. Worked with data engineers to implement real-time data monitoring.
4. Retrained the model with up-to-date data and adjusted hyperparameters to improve generalization.

This fix reduced false positives by 25%, improved fraud detection accuracy, and ultimately saved $500M in operational savings across multiple initiatives. This experience taught me the importance of continuous model monitoring and proactively accounting for data drift. Now, I always incorporate data validation pipelines and feedback loops into AI deployments to prevent similar issues.

let's consider two vectors y and y_bis, the root mean square error between these two vectors is equal to the L2-norm between y and y_bis. in another words, let's consider z = y - y_bis. the root mean squared error corresponds to the square root of the sum square element of z.

I will code it in Python : 

first thing to have in mind is to code the sqrt function to compute square root of a number. for this purpose we use newton method.

def sqrt(x: float, epsilon: float = 1e-10) -> float:

    if x < 0:

        raise ValueError("Cannot compute square root of negative number.")

    if x == 0 or x == 1:

        return x

    guess = x / 2.0

    while abs(guess * guess - x) > epsilon:

        guess = (guess + x / guess) / 2.0

    return guess

now that we have the sqrt function we can define the root mean squared error function as :

def rmse(y: List[float], y_bis:List[float]) -> float:

    return sqrt(sum((y-y_bis)**2))

sum() is a native function in python.

When I joined my new team , I took lead on a new project which was complex in nature, had many unknowns and our team had to work on new framework, techstack, data models. The team was following agile methodology and will start with the work in first sprint. there was no concept of zero sprint or story grooming prior to the start of first sprint.

with the existing approach I envisioned some issues:

1. Because the number of unknown were high, the chances of doing wrong estimations for development were high.

2. There were high chances that the prioritisation is not right and we might have to revisit the features again and again.

3.  We might find out about unknowns while we are in the development process and that might cause delays.

To overcome these issues, I suggested some changes in our approach:

1. Increase the planning for one more release cycle and group the features as now, next and later.

2. For the features planned in the now group , prioritise them for the next release and complete the research and PRDs for them one sprint early. That means dev and QA will have 2 weeks extra to go through the stories, find out the dependencies and share the feedback. This helped in identifying if we need POCs for any feature or we can right away start developing them. Or if further modularisation is required.

3. I suggested to track the features related to new project in a different group, because to begin with the pace of the delivery will be slow.

When we begin, we took one complete release only in the RnD and understanding the new framework and doing small POCs to validate our understanding.

From second release, when we as a team were ready with good understanding on the expectations and other tech dependencies we started delivering and we could complete 2 full flows from end to end.

We had very less amount of spillover and could complete almost everything in the backlog, which was a big issue in earlier releases.

Bootstrapping is a statistical resampling method used to estimate the distribution of a sample statistic (like the mean, median, variance) by repeatedly sampling from the original dataset with replacement. The primary idea is to generate "new" samples (called bootstrap samples) by randomly selecting data points from the original sample, allowing some points to be selected multiple times while others may not be selected at all.

Steps of Bootstrapping:

1. Original sample: Start with a dataset of size n.
2. Resampling: Generate multiple new datasets (bootstrap samples) of the same size n by sampling with replacement from the original dataset.
3. Statistic calculation: For each bootstrap sample, calculate the statistic of interest (e.g., the mean).
4. Aggregation: After many resamplings (typically thousands), aggregate these statistics to estimate properties like confidence intervals, standard errors, or the distribution of the statistic.

Efficacy in Augmenting Sample Size:

While bootstrapping doesn’t actually create new, independent data, it is effective at enhancing statistical insights from small samples by simulating variability and giving a better approximation of the underlying population’s distribution. Its efficacy is most pronounced when:

• Small samples: Bootstrapping is especially useful for datasets where traditional parametric methods may not be applicable due to the small sample size or assumptions (like normality).
• Non-parametric nature: It does not require assumptions about the distribution of the data, making it versatile.
• Uncertainty Estimation: It helps estimate confidence intervals, standard errors, and biases for small samples when direct analytical solutions are difficult.

However, since bootstrapping is based on the assumption that the original sample is representative of the population, its effectiveness can be limited when the original sample is biased or unrepresentative. It’s not a substitute for truly increasing the sample size but is a powerful technique for making the most of available data.

with p the general probability of rain in Dublin, P(rainy | 3 friends say it's raining) = 8/(7+1/p)

At Auctane, I had a Sr. Engineer who was very harsh with Jr. Engineers in giving feedback. This led to Jr. Eng afraid of approaching with advice and team morale go down and productivity was down by 10%. I had to coach this teammate to not be hostile and instead approach people with Empathy and show ledership in asking the right questions and help them grown in their carrer as well - with this change in attitude, I saw that the team gel much better and the productivity increased overtime.