Affine Analytics Prompt Engineer Interview: Rounds, Questions and Preparation

What the job involves

Prompt Engineers in AI and deeptech companies design, test, and refine prompts that make large language models produce reliable, useful outputs for products like analytics copilots, document Q&A, and workflow automation. Day to day you write prompt templates, run evaluation experiments, compare model versions, and work with data, product, and engineering teams to ship AI features. At fast-growing firms like Affine Analytics, you often own prompt quality for a specific client use case and iterate quickly based on user feedback.

Role and technical knowledge

1. 1

   What is the difference between zero-shot, few-shot, and chain-of-thought prompting, and when would you use each?

   Define each pattern clearly, then give a concrete use case like classification, extraction, or multi-step reasoning. Mention that few-shot helps with format and tone while chain-of-thought helps with arithmetic or logic.

2. 2

   How do temperature and top-p affect LLM output, and how would you set them for a customer-facing analytics assistant?

   Explain that low temperature and low top-p make outputs more deterministic, which is usually safer for factual or structured answers. Mention that you would test a few settings on a small eval set before deciding.

3. 3

   What is retrieval-augmented generation and why is it important for enterprise analytics use cases?

   Describe RAG as fetching relevant documents or database rows and inserting them into the prompt. Explain that it reduces hallucination and lets the model use private or up-to-date data without retraining.

4. 4

   How would you force an LLM to return clean JSON for a downstream analytics pipeline?

   Mention using JSON mode or function calling if the API supports it, giving a schema in the prompt, adding a validation step in Python, and retrying or fixing outputs that fail parsing.

Practical scenarios

1. 5

   A client says the AI assistant is giving wrong numbers for monthly sales. How do you debug the prompt and the pipeline?

   Break the problem into retrieval, prompt, and model layers. Check whether the right data was fetched, whether the prompt asks for calculations or just formatting, and whether a deterministic tool like SQL should handle the math instead.

2. 6

   You need to build an evaluation set for a prompt that summarizes long analyst reports. What does your eval set look like and what metrics do you track?

   Describe 30 to 100 representative examples with expected key points. Track factual consistency, coverage of key points, hallucination rate, length, and latency. Mention human review for a sample and automated checks where possible.

3. 7

   How would you reduce the cost and latency of a prompt that currently sends a 20-page document to the model for every user query?

   Suggest chunking and retrieving only relevant sections, summarizing once and caching the summary, trimming few-shot examples, and testing a smaller model for simpler queries.

4. 8

   A user tries to trick the assistant into revealing its system prompt. How do you design guardrails?

   Mention separating system instructions from user input, adding a policy check before and after the model call, refusing unsafe requests, and testing with known injection patterns. Avoid claiming any single technique is foolproof.

Behavioural and HR

1. 9

   Why do you want to work as a Prompt Engineer at a fast-growing AI and analytics company rather than a large tech firm?

   Show that you understand the pace, breadth, and client-facing nature of smaller AI companies. Connect it to your interest in shipping quickly and learning across data, product, and engineering.

2. 10

   Tell me about a time you had to learn a new tool or concept quickly for a project or hackathon.

   Use a specific example with a clear situation, action, and result. Highlight how you found reliable resources, built a small prototype, and measured whether it worked.

3. 11

   Describe a situation where your first solution did not work and you had to iterate. How did you decide what to change?

   Focus on a structured debugging approach: define the failure, form a hypothesis, change one variable, and measure again. Show comfort with iteration rather than defending the first attempt.

4. 12

   How do you explain a technical prompt-engineering trade-off to a non-technical client or teammate?

   Give an example like accuracy versus cost or speed versus creativity. Use plain language, a simple analogy, and a clear recommendation with the trade-off stated honestly.

Take-home and portfolio review

1. 13

   Walk me through a prompt you wrote for a project, hackathon, or personal experiment. What problem did it solve?

   Describe the task, the initial prompt, what failed, how you changed it, and what metric improved. Strong answers show a before-and-after comparison, not just a final prompt.

2. 14

   How do you document a prompt so another engineer can maintain it six months later?

   Mention versioning the prompt, stating the model and settings used, listing known failure cases, and including the evaluation results. Treat the prompt like code with comments and a changelog.

3. 15

   If you had to choose between a prompt that is 95 percent accurate but slow and one that is 85 percent accurate but fast for a live analytics dashboard, what would you do?

   Ask about user expectations and whether the answer is cached or interactive. Suggest a hybrid: fast path for common queries and slower path for complex ones, with accuracy monitored per path.

4. 16

   What is one limitation of current LLMs that you think prompt engineers should always keep in mind?

   Pick a real limitation like hallucination, sensitivity to phrasing, or poor arithmetic. Explain how it affects prompt design and what mitigation you would use in production.