This document contains the exact, high-weightage interview questions asked at Google, Meta, Microsoft, Accenture, and Infosys for AI, ML, and Cloud Security roles.
Because you built the Zero-Trust GenAI Architect, you can answer all of these questions by pointing directly to your architecture.
Focus: Generative AI, RAG, LLMs, Langchain, and AI Automation.
Q1: "How do you prevent an LLM from hallucinating when dealing with critical company infrastructure?"
The 1-Crore Answer:
"In my Zero-Trust Architect project, I could not allow the LLM to 'guess' how to fix an AWS server. To prevent hallucination, I implemented a RAG (Retrieval-Augmented Generation) architecture. Before the LLM generates any Terraform patch, it first queries a local Vector Database containing strict Enterprise Security Policies. I inject the exact AWS/Azure security baseline directly into the prompt context. This mathematically anchors the LLM's output to the company's approved standards, reducing hallucination to near zero."
The 1-Crore Answer:
"Yes, my entire architecture is based on an asynchronous Multi-Agent framework. Instead of a single script, I decoupled the logic. I used Langchain to build Agent 3 (The Auto-Patcher). It takes the vulnerability detected by the ML Agent, retrieves the policy via RAG, and generates the exact JSON/Terraform payload required to fix the vulnerability dynamically."
Focus: Graph ML, Scalability, Anomaly Detection, Scikit-Learn.
Q3: "Standard Machine Learning models struggle with Cloud Security because the data is just JSON configurations. How did you apply ML to this problem?"
The 1-Crore Answer:
"Standard tabular models fail here. In my project, I applied Graph Machine Learning. Cloud infrastructure is not a flat table; it's a web of connected resources (e.g., an S3 bucket attached to an IAM role attached to a VPC). I used Python's
NetworkXto parse the AWS JSON and represent the entire cloud as a Directed Graph. I then wrote ML algorithms to traverse the edges and detect 'Toxic Combinations'—for example, a node withpublic_read=Truedirectly connected to a node withadmin_access=True."
The 1-Crore Answer:
"For Cost-Effective Scale. If I sent the entire AWS infrastructure state to OpenAI every 5 seconds, it would cost the company millions in API tokens. I used traditional ML (Graph Traversal and Isolation Forests) as a 'Filter'. The ML runs locally on CPU for free and drops 99% of normal configurations. The expensive LLM is only triggered when the ML definitively flags a Toxic Combination."
Focus: AWS/Azure, Zero-Trust, IAM, Terraform, DevSecOps.
The 1-Crore Answer:
"Zero-Trust means 'Never trust, always verify', even inside the internal network. In my project, I automated Zero-Trust posture management. When a developer deploys an Azure VM with Port 22 open to the internet, my system detects the misconfiguration. But instead of just sending an alert, the system uses Generative AI to automatically rewrite the Azure SDK Network Security Group (NSG) rules to restrict access to a specific /8 subnet. It actively enforces Least Privilege without human intervention."
Q6: "How do you manage security across multiple cloud providers (AWS and Azure) without writing two separate platforms?"
The 1-Crore Answer:
"I built a Unified Threat Schema using FastAPI and Pydantic. Whether the log comes from AWS EventBridge or Azure Sentinel, my API gateway standardizes the JSON into a generic Graph Node format. This allowed my Graph ML and RAG engine to be 'Cloud Agnostic'. I don't maintain two codebases; I maintain one AI engine that understands both AWS and Azure syntaxes."
When interviewing at Accenture or Infosys, they will be highly impressed by the Multi-Cloud (AWS + Azure) aspect. When interviewing at Google, Meta, or Microsoft, they will dive deep into the RAG Vector Database and Graph ML architecture. Use the answers above to dominate the technical rounds.