A better way to prepare.

Interview-ready answer

A large language model is a neural network trained on large collections of text to predict likely next tokens. When given a prompt, it converts the text into tokens, uses a Transformer to process their context, and generates a response one token at a time. It learns patterns rather than checking every statement against a trusted source, so its answers can sound plausible and still be wrong.

Interview-ready answer

Retrieval-Augmented Generation (RAG) retrieves relevant material from an external knowledge source and supplies it to a language model alongside the user’s question. A typical system chunks and indexes documents, retrieves likely matches at query time, and asks the model to answer using that context. RAG is useful when answers should reflect private, changing, or cited information, but retrieval quality and source quality still limit the result.

Interview-ready answer

Supervised learning learns from examples paired with target labels, such as emails marked spam or not spam. Unsupervised learning looks for structure in data without provided target labels, such as grouping similar customers. Reinforcement learning learns a policy by taking actions and receiving rewards or penalties over time. The key distinction is the learning signal, not whether an algorithm is “intelligent.”

Interview-ready answer

Overfitting occurs when a model fits details or noise in its training data so closely that it performs poorly on new examples. A common sign is strong training performance but weaker validation performance. You can reduce the risk with representative data, appropriate model complexity, regularization, early stopping, and careful validation. The goal is generalization, not simply minimizing training error.

Interview-ready answer

MLOps is the set of practices and tools for reliably developing, deploying, and maintaining machine-learning systems. It extends software delivery with data and model concerns such as reproducible training, validation, model and dataset versioning, monitoring, and retraining decisions. MLOps matters because model behavior depends on data and can change over time, so deployment is the start of ongoing operations, not the end.

Interview-ready answer

Model drift is a decline in predictive usefulness as the relationship between production data and the task changes. Data drift describes a change in input distributions; concept drift describes a change in how inputs relate to outcomes. Teams detect possible drift by monitoring inputs and delayed performance, investigate the cause, and validate a response such as updating data or retraining before release.

Interview-ready answer

A list is an ordered, mutable sequence that allows duplicates. A tuple is an ordered sequence that cannot be reassigned item by item. A set stores unique hashable values and is useful for membership and set operations. A dictionary maps unique hashable keys to values. Choose based on whether you need order, mutation, uniqueness, or key-based lookup.

Interview-ready answer

Scan the string once while storing characters already seen in a set. If the current character is already present, the string contains a duplicate; otherwise add it and continue. This takes O(n) expected time and O(k) extra space, where n is the string length and k is the number of distinct characters. Clarify whether comparison is case-sensitive and whether spaces count.

Interview-ready answer

ETL extracts data, transforms it before loading, and writes the prepared result to a destination. ELT extracts and loads raw data first, then transforms it inside the destination platform. ETL can fit systems that need curated data before storage; ELT often uses scalable warehouses to retain raw data and transform it later. The choice depends on privacy, compute, latency, and governance needs.

Interview-ready answer

A data pipeline moves data from sources through ingestion and transformation to storage or downstream consumers. A reliable design defines data contracts, validates inputs, makes retries safe through idempotent writes, handles duplicates and late records, and tracks each run. Orchestration manages dependencies; monitoring and alerting expose failures and quality issues. Recovery and backfills should be designed rather than improvised after an outage.

Interview-ready answer

I would separate the write path that creates a short code from the read-heavy redirect path. The API validates the destination, generates a unique code, and stores the mapping with expiry and ownership metadata. Redirects use a cache in front of durable storage, with careful handling for missing or expired codes. I would clarify traffic and reliability requirements first, then address abuse controls, analytics, and cache consistency.

Interview-ready answer

A rate limiter decides whether a request is allowed under a policy such as N requests per identity per time interval. I would place it at an API gateway or service boundary, choose an algorithm such as token bucket based on burst behavior, and use atomic state updates when limits are shared across instances. The design must define identity, response headers, storage failures, and whether temporary inconsistency is acceptable.