Company Interview Questions · IT Services

Infosys: GenAI Interview Experience

❓ 11 questions🏢 Infosys

Infosys Gen AI Interview Experience (2026-2027) — Complete Transcribed Questions & Answers. Real questions from the interview, with detailed answers. Practice mode hides every answer: say your answer first, press Show answer, then rate yourself. Use 🔁 Review again to build your revision list and 🎲 Random question for a quick drill.

★ Must-know 145📝 Assignments

01Introduction, Project Discussion & Agentic Workflows

Q1
Tell me about yourself and what exactly are you working on currently?
  • Project Overview: Discussed hands-on experience building production-grade Generative AI applications, focusing on architectural design, RAG implementations, and agentic workflows.
  • Individual Contribution: Explained personal ownership in designing the orchestration layer, integrating external tools, and managing backend execution flows.
Q2
You mentioned Agentic AI. What exactly is Agentic AI in your application?
  • Definition & Core Logic: An AI agent goes beyond simple single-step text generation by using an LLM as a central reasoning engine to evaluate inputs, determine required actions, invoke external tools or APIs, receive feedback, and continue execution loops autonomously until the task is complete.
Q3
Who decides which tool to call, and what if the model selects the wrong tool?
  • Tool Orchestration: The model selects appropriate tools based on the task prompt and available tool descriptions, while the application's orchestration layer strictly controls which tools are exposed and what permissions they hold.
  • Handling Incorrect Tool Selection (Production Scenarios):
  • Use clear, explicit tool descriptions and parameter schemas (e.g., Pydantic).
  • Implement restricted tool access and strict input/output validation.
  • Build robust fallback handling and loop limits so the agent fails gracefully or asks for clarification rather than cascading errors.

02Model Context Protocol (MCP)

Q4
Have you worked with MCP? Explain MCP in your own words.
  • Definition: The Model Context Protocol (MCP) is an open standard that provides a standardized way for AI applications (model hosts) to connect with external tools, data sources, and context providers.
  • Components: MCP servers expose capabilities like tools, resources, and prompts, which an MCP host can dynamically discover and connect to.
Q5
What is the difference between normal function calling and MCP?
  • Traditional Function Calling: Implemented directly and custom-built between a specific monolithic application and its hardcoded functions.
  • Model Context Protocol (MCP): Provides a universal, standardized protocol layer. Instead of custom point-to-point integrations for every tool and application, MCP decouples the client from external capabilities, allowing any MCP-compatible host to dynamically discover and plug into various MCP servers seamlessly.

03RAG vs. Fine-Tuning

Q6
When would you choose fine-tuning over RAG, and can you use both?
  • RAG vs. Fine-Tuning: They solve different problems and are not mutually exclusive competitors.
  • Use RAG: When the model needs access to external, private, and frequently changing enterprise knowledge without retraining.

  • Use Fine-Tuning: When you need to modify the model's fundamental behavior, tone, style, formatting rules, or specialized domain-specific task capabilities.

  • Combining Both: Yes, you can use both simultaneously—for example, deploying a fine-tuned model optimized for domain tone that is also grounded via real-time context retrieved through a RAG pipeline.

04LLM & RAG Evaluation

Q7
How do you know your Gen AI application is actually getting better? (Offline vs. Production Evaluation)
  • Multi-Level Evaluation: Relying solely on user satisfaction ("users are happy") is insufficient. Evaluation must be systematic:
  • Retrieval Layer: Separately evaluate the retriever using precision@K, recall@K, and ranking metrics.

  • Generation Layer: Evaluate generated outputs against retrieved context using metrics like faithfulness, relevance, and correctness.

  • Handling Discrepancies (Offline Benchmarks vs. Production Realities):

  • If offline evaluation scores look great but users still complain, it means benchmark datasets do not reflect real-world user distribution.
  • Resolution: Capture actual production failure traces, difficult queries, edge cases, and user feedback, and continuously feed those real-world examples back into the evaluation test suite.

05Python Engineering & Backend Fundamentals

Q8
Why would you use async programming in an AI application? Suppose your application has to call three independent APIs—would you call them sequentially or concurrently?
  • Async Programming: Essential in AI backends to handle I/O-bound operations efficiently without blocking the event loop while waiting on external network calls.
  • Concurrent Execution: For three independent API calls (e.g., fetching user data, querying a vector DB, and checking order status), use concurrent/asynchronous execution (e.g., asyncio.gather) instead of sequential calls to drastically reduce overall latency and waiting time.
Q9
What is the difference between concurrency and parallelism, and where does the Python GIL become relevant?
  • Concurrency vs. Parallelism: Concurrency is about dealing with lots of things at once (task switching, ideal for I/O-bound tasks), while parallelism is about doing lots of things at the exact same instant (true multi-core execution, required for CPU-bound math like embeddings or text chunking).
  • Python GIL (Global Interpreter Lock): A mutex in CPython that prevents multiple native threads from executing Python bytecodes at once. It becomes relevant in multi-threaded applications because it restricts CPU-bound parallel processing, necessitating the use of multiprocessing for heavy CPU workloads.

06Advanced Agentic Scenarios & Production Guardrails

Q10
Scenario: Imagine your agent has access to 10 tools. The user asks a simple question, but the agent starts calling 5 different tools. How would you control this?
  • Tool Selection Controls: Having access to tools does not mean the agent should use all of them blindly.
  • Implement strict tool selection constraints, routing logic, and system prompt instructions that guide the agent on when a tool is genuinely necessary.
  • Use deterministic routing for known, repetitive workflows to bypass unnecessary LLM reasoning steps.
  • Set strict iteration limits and timeout thresholds on tool calls to prevent runaway loops.
Q11
Would you completely trust the LLM to perform a sensitive action like deleting a customer record?
  • No. Unrestricted LLM power over production systems introduces severe operational and security risks.
  • Production Safeguards: Enforce application-level authorization, role-based access control (RBAC), and mandatory Human-in-the-Loop (HITL) approvals before any state-changing, destructive, or sensitive action (like data deletion or financial transactions) is executed.
Created by Ajeetkumar