AI/ML and generative AI course with interview preparation, 1‑on‑1 with a working AI engineer.

Overfitting is when a model learns the training data too closely, including its noise, and does poorly on new data. You see it when validation results are much worse than training results. Fixes include more data, simpler models, regularisation, cross-validation and early stopping.

Precision is how many predicted positives were correct; recall is how many actual positives you found. Fraud or disease screening usually cares more about recall, spam filtering more about precision. The F1 score balances the two.

A token is a chunk of text, often part of a word, that an LLM reads and generates. The context window is the maximum number of tokens the model can consider at once, including your prompt and its answer. It limits how much information you can give the model in one request.

Retrieval-augmented generation finds relevant documents and gives them to the model with the question, so answers are grounded in your data. It’s usually cheaper and easier to keep up to date than fine-tuning, which changes a model’s behaviour or style rather than adding fresh knowledge.

It depends on the documents and the questions. Chunks that are too small lose context; chunks that are too large make retrieval noisy. Split along natural boundaries such as headings, use some overlap, keep metadata, and test chunk sizes against real questions instead of guessing.

Separate retrieval from generation. Check whether the right chunks were retrieved at all; if not, fix chunking, embeddings or filters, or add re-ranking. If they were, look at the prompt, whether it allows “I don’t know”, and how the context is ordered. Interviewers want to hear how you measured it.

Build a set of real questions with expected answers and measure correctness, groundedness in the sources and relevance, using automatic checks, LLM-as-judge scoring and human review. Track cost and latency too, and re-run the set every time you change the prompt or model.

Choose metrics that reflect it, such as precision, recall or PR-AUC rather than accuracy. Then try resampling, class weights or threshold tuning, and validate on data with the real-world balance.

Ingestion and chunking with access control, embeddings and a vector store, retrieval with re-ranking, a prompt that cites sources, then evaluation, monitoring and a feedback loop. Interviewers probe permissions, how fresh the data is, cost, and what happens when the answer isn’t in the documents.

Use smaller models where they’re good enough, cache repeated requests, trim prompts and retrieved context, stream responses and send only hard requests to larger models. Measure before and after; senior interviews want numbers from something you’ve shipped.

When you need a consistent style, format or behaviour that prompting can’t deliver reliably, or a smaller, cheaper model for a narrow task. For new or changing knowledge, RAG is usually better. Fine-tuning needs good data and an evaluation set before you start.

Watch incoming data for drift, track prediction quality where you can get labels, monitor latency, errors and cost, and set alerts. For LLM apps, review a sample of real conversations and track user feedback.