To succeed in the technical rounds, you must demonstrate mastery across several core domains. Infosys interviewers will evaluate your practical engineering experience rather than just theoretical knowledge.
Large Language Models & Tokenization
Understanding the underlying mechanics of LLMs is critical for optimizing performance and cost in enterprise environments.
You must be prepared to discuss how text is transformed into a format that models can process and how this affects overall system behavior.
Be ready to go over:
- Tokenization Mechanics – How Byte-Pair Encoding (BPE) or WordPiece algorithms segment text, and why certain characters or languages consume more tokens.
- Context Window Management – Strategies for handling long inputs, including truncation, sliding windows, and prompt compression techniques.
- Model Selection & Trade-offs – When to utilize proprietary models (like GPT-4) versus hosting open-source alternatives (like LLaMA or Mistral) based on latency, privacy, and cost.
- Advanced concepts (less common) – Fine-tuning methodologies (such as LoRA and QLoRA), understanding KV-caching, and optimizing inference speeds using quantization (e.g., AWQ, GPTQ).
Example questions or scenarios:
- "How would you optimize a prompt template to minimize token usage while maintaining high accuracy for a classification task?"
- "Explain how a tokenizer's handling of special characters can affect the performance of an LLM designed to parse code or structured JSON."
Retrieval-Augmented Generation (RAG) & Chunking
Since most enterprise data is proprietary, building high-quality RAG pipelines is one of the most common tasks for a GenAI Engineer at Infosys.
You must demonstrate a comprehensive understanding of how to ingest, index, retrieve, and synthesize information effectively.
Be ready to go over:
- Chunking Strategies – The pros and cons of fixed-character chunking, overlapping chunks, document-based chunking, and semantic chunking based on sentence similarity.
- Embedding Models – How to select and evaluate embedding models for specific domains (e.g., financial, legal, or medical text).
- Vector Search & Indexing – Understanding hierarchical navigable small world (HNSW) graphs, inverted file index (IVF), and how to scale vector search in production.
- Advanced concepts (less common) – Implementing self-RAG, query rewriting, metadata filtering, and utilizing cross-encoder re-ranking models to improve precision.
Example questions or scenarios:
- "Walk me through how you would design a RAG system for a client with millions of multi-page PDF documents containing complex tables and images."
- "If your RAG system is consistently retrieving irrelevant information, what debugging steps would you take to isolate the issue?"
Agentic AI & Orchestration
Modern AI systems are transitioning from static pipelines to dynamic, autonomous agents that can plan, use tools, and collaborate.
Interviewers will evaluate your ability to design systems that can reason and execute multi-step tasks.
Be ready to go over:
- Tool-Calling & Function Binding – How to define schemas that allow an LLM to reliably generate arguments for external API calls.
- State and Memory Management – How to maintain short-term and long-term memory across complex, multi-turn conversations and agent loops.
- Framework Proficiency – Your practical experience with orchestration libraries such as LangChain, LangGraph, LlamaIndex, or CrewAI.
- Advanced concepts (less common) – Implementing the Model Context Protocol (MCP), managing multi-agent negotiation, and preventing infinite loops in autonomous execution.
Example questions or scenarios:
- "Design an agentic system that can take a user's natural language request, query a SQL database, format the result into a CSV, and email it to a stakeholder."
- "How do you ensure that an AI agent does not execute harmful or unintended tool calls when processing untrusted user input?"
