Interviews For AI Engineer Role (2025 H2)

To See All Interview Preparation Articles: Index For Interviews Preparation

1: Ericsson at Accenture

Q: What do you know about MCP?

Q: What is the difference between an MCP server and an API?

Q: Can we setup asynchronous connection with an MCP server? If yes, how?

### ### ### ### ### ### ### ### ### ### ### ### ### ### ### 

2: Macquarie (Australian Financial Firm)

Q: Talk about your overall experience. 

Q: Talk about your current/last project.

Q: What's prompting you to leave Accenture? What is that trigger?

Q: What do you know about Macquarie?

### ### ### ### ### ### ### ### ### ### ### ### ### ### ### 

3: abc@relx.com at a Legal Firm

Q: Can you explain the architecture of your current / last Agentic AI project?

Q: How would you evaluate a custom model that you have trained for classification purpose?

Q: Why and how did you choose framework for your Agentic application between LangGraph, LangChain, CrewAI, AutoGen, AI Refinery?

Q: Can you build a simple Agentic workflow with one agent and one or two tools (getWeather, getTraffic)?

4: Barclay's @ Accenture

Q1: I was asked about my AI/ML journey.
- Answer: Start with your latest project:
# Content Optimization platform in Ed-tech domain
# Then on top of that content, building mobile apps later donated to NGO in Educational cuase as part of CSR initiative of Accenture.

# Built three apps using vibe coding using Agentic Coding platforms such as Windsurf and GitHub Copilot.

In simple terms, I was generating content for the Ed-tech mobile apps first, seeding them in a SQLite database (so that the user has a local and offline access to the data), then vibe-coding the mobile apps.

... ... ... ... ...

Q2: Challenges faced in this project were:
- content relevance 
- grammatical correctness 
- complexity of content for different age groups

Q3: Explain your last project -- AI Over BI
- what it did
- how it worked
- talk a bit on the architecture

Q4: So you used LLM-As-A-Judge to reflect on the query generated by an LLM and produce some feedback. The question is: how do you know LLM-As-a-Judge is not hallucinating itself and that it is providing the desired feedback?

Q5: How does your application come out of generate-reflect-regenerate loop?
Answer: Two techniques to come out of loop:
1. Max iterations
2. Consistency check between last and second-last generated query.

Q6: How do you know that the two queries are consistent? What metric are you employing there?

Q7: Do you have a hands-on experience with 'procuring and setting up cloud resources' to 'building solution' to 'deployment on the cloud'?

Q8: If you observe that there is a latency issue with your RAG solution, what can you do about it?

Answer:
1. Trying out a faster LLM
2. Caching
3. Collecting component wise traces for checking which component is taking maximum time (viz usually the LLM but the traces will tell which is the second highest)

Q9: Which chunking strategy did you use and why?

Q10: Why didn't you use Semantic Chunking?

Q11: How did you prevent or reduce hallucinations in your RAG agent?

----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- 

5: IBM (Jan 2026)

Q1: Question on Pandas library around the functions such as merge(), groupby() and corr().

Q2: Have you built AI agents from scratch? If yes, using which framework?

Q3: Distinguish between AI Agent and AI Assistant.

Q4: Why does one-shot works better than zero-shot prompts?

Q5: Give an example where zero shot prompting seems to work just fine.

Q6: What is the structure of your current team?

Q7: Are you in a client facing role?

Q8: Describe a situation where you had to handle a client escalation?

Q9: Why are you leaving Accenture?

Q10: Why do you want to join IBM?

Tags: Interview Preparation,