New opportunity

Senior AI Data Engineer

EXL Talent Acquisition Team · Gurugram, India

About this role

EXL Talent Acquisition Team seeks a Senior AI Data Engineer to develop production-grade LLM applications for business use cases. The position is located in Gurugram, Haryana, India.

Responsibilities

Design single- and multi-agent applications and build end-to-end RAG pipelines covering ingestion, embeddings, retrieval, orchestration and response synthesis.
Implement prompt chaining, tool and function calling, and structured outputs. Develop GenAI APIs and services using tools such as FastAPI, Flask or Streamlit, and integrate solutions with enterprise systems, data platforms and workflows.
Apply guardrails and evaluation frameworks to improve response quality, reduce hallucinations and support responsible AI use. Work with Data Engineering and MLOps teams on pipelines, deployment, monitoring and scaling; contribute reusable components, documentation and engineering practices.

Required qualifications

6–9 years of total experience, including 1–3+ years of hands-on GenAI or LLM application development for production use cases.
Hands-on experience with LLMs such as Claude or OpenAI, RAG and retrieval optimization, GPT and agentic AI implementation, and frameworks such as LangChain or LangGraph. Understand agent orchestration, tool-calling architectures, and LLM limitations, evaluation and optimization.
Strong production-grade Python and PySpark engineering, proven API integration, analysis of large data volumes, and data engineering integration with Fabric, Azure Databricks or Snowflake. Exposure to Azure, AWS or GCP, SQL, containers, CI/CD and monitoring is expected.
Prior experience in at least one of data engineering, including ETL/ELT and pipeline orchestration; the data science or ML lifecycle, especially NLP; or analytics engineering and data products.

Preferred qualifications

Fine-tuning with LoRA or PEFT, or prompt tuning; enterprise GenAI security and privacy practices such as data masking, access control and compliance; familiarity with Azure OpenAI, Azure AI Search or Fabric; and exposure to agentic coding tools such as Claude Code.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

Generative AILarge Language ModelsAgentic AIRAGPrompt engineeringPrompt chainingTool callingLangChainLangGraphPythonPySparkFastAPIFlaskStreamlitAPI integrationData engineeringETLELTData analysisSQLMicrosoft FabricAzure DatabricksSnowflakeAzureAWSGoogle Cloud PlatformContainersCI/CDMonitoringMLOps

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

End-to-end RAG quality
The role builds and optimises ingestion, embeddings, retrieval, orchestration and response synthesis. Prepare to explain how you would diagnose retrieval failures and improve answer quality across the full pipeline.
Agentic application design
The posting calls for single- and multi-agent applications, prompt chaining, tool/function calling and structured outputs. Prepare to justify orchestration choices and explain how tool results become usable responses.
Production services and data integration
The role develops GenAI APIs and services and connects them to enterprise systems, workflows and data platforms. Production-grade Python/PySpark, API integration, large-volume data handling and Fabric, Azure Databricks or Snowflake integration are core requirements. Prepare an end-to-end design explanation rather than a tool list.
Evaluation and responsible AI
The role applies guardrails and evaluation frameworks to improve responses, reduce hallucinations and support responsible use. Prepare to describe how you would test limitations, measure changes and decide whether an application is ready to deploy.
Deployment and reusable engineering
The role works with Data Engineering and MLOps teams on pipelines, deployment, monitoring and scaling, while contributing reusable components and documentation. Prepare to explain ownership boundaries, operational signals and how a component could be reused.

Likely challenges to think through

Retrieval quality versus pipeline complexity
Likely challenge, not a stated employer problem: improving answers may require changes to ingestion or retrieval, not just prompts, while each change adds work to evaluate. Prepare a stepwise diagnosis that separates missing evidence, poor ranking and response-synthesis errors.
Agent flexibility versus control
Likely challenge, not a stated employer problem: more tool calls or agents can handle complex tasks but introduce failure paths and less predictable outputs. Prepare criteria for choosing a single agent, multiple agents or a simpler chain, including how you would validate tool calls and structured outputs.
Scaling while preserving response quality
Likely challenge, not a stated employer problem: enterprise integration and higher data volumes can make it harder to maintain observable, reliable responses. Prepare to explain how deployment monitoring and evaluation would detect regressions, and how you would coordinate fixes with Data Engineering and MLOps.

Practice your answers

Technical practice

How would you design an end-to-end RAG pipeline for a business use case?

What to coverSuggested answer focus: Walk through ingestion, embeddings, retrieval, orchestration, and response synthesis, all named in the posting. Explain how you would keep source data current and trace an answer back to retrieved content.

How would you diagnose and improve poor retrieval quality?

What to coverSuggested answer focus: Separate ingestion or embedding problems from retrieval-ranking problems, then test changes against representative queries. Discuss the trade-off between retrieving enough relevant context and adding distracting material.

When would you use a single agent versus multiple agents?

What to coverSuggested answer focus: Explain how task boundaries, tool access, and orchestration complexity affect the choice. Describe how you would control handoffs and handle tool failures rather than assuming more agents produce better results.

How would you implement prompt chaining, tool calling, and structured outputs?

What to coverSuggested answer focus: Describe each step's inputs and outputs, validation of tool arguments and returned data, and recovery from malformed responses. Explain when a simpler prompt would be preferable to a chain.

How would you build a production-grade GenAI API that integrates with enterprise systems?

What to coverSuggested answer focus: Discuss a Python service using a framework named in the posting, such as FastAPI or Flask, and define its request, response, and integration boundaries. Address error handling, access control, and the trade-offs introduced by downstream dependencies.

How would you evaluate an LLM application and reduce hallucinations?

What to coverSuggested answer focus: Describe a representative evaluation set and measures for answer quality, grounding, and failure cases. Explain how guardrails and retrieval improvements would be tested together, including cases where the system should decline to answer.

How would you prepare large-volume data for an LLM application?

What to coverSuggested answer focus: Connect Python or PySpark processing with ETL/ELT, data analysis, and integration with a platform named in the posting, such as Fabric, Azure Databricks, or Snowflake. Discuss data quality, repeatable pipeline runs, and the effect of upstream changes on retrieval.

How would you deploy, monitor, and scale an LLM-backed service?

What to coverSuggested answer focus: Cover containers, CI/CD, and monitoring in the context of collaborating with MLOps. Explain what you would observe across service health, latency, cost, and response quality, and how those signals would guide scaling decisions.

What security and privacy controls would you apply to enterprise GenAI data?

What to coverSuggested answer focus: This is a preferred area in the posting; discuss data masking, access control, and compliance considerations without assuming a particular policy. Explain how controls should apply to both retrieved content and tool-mediated actions.

When would you choose prompt tuning over fine-tuning techniques such as LoRA or PEFT?

What to coverSuggested answer focus: These are preferred, not mandatory, topics in the posting. Compare expected quality gains with data needs, evaluation effort, maintenance, and deployment complexity before choosing an approach.

Behavioural practice

Tell me about coordinating a delivery with Data Engineering and MLOps teams.

What to coverSuggested practice question: Give a real situation, your actions to agree on pipeline, deployment, or monitoring responsibilities, and the result. Distinguish your contribution from the teams' contributions.

Describe a time you investigated a quality problem in a production AI or data system.

What to coverSuggested practice question: State the symptoms and impact, how you isolated the cause and validated a fix, and the result. If the example involves an LLM, explain what it taught you about evaluation or guardrails.

Tell me about integrating a technical solution into an existing business workflow.

What to coverSuggested practice question: Describe the workflow constraint, how you worked with its stakeholders and addressed integration trade-offs, and the result. Keep the example grounded in work you actually did rather than assuming a specific EXL workflow.

Describe a reusable component or documentation you contributed to a team.

What to coverSuggested practice question: Explain the repeated problem, your design and documentation choices, and the resulting effect on other engineers. Clarify how you checked that the component was useful and maintainable.

Tell me about making a trade-off between AI response quality and operational constraints.

What to coverSuggested practice question: Describe the situation, the alternatives you evaluated, your decision, and its observed result. Relevant constraints could include latency, scale, or responsible use; discuss only those that applied to your example.

Sharpen your knowledge

Topics to prepare

RAG architecture and retrieval optimisation
The posting specifies ingestion, embeddings, retrieval, orchestration and response synthesis. Revision exercise: diagram that flow and explain how you would diagnose a relevant document being missed versus a correct document producing a poor answer.
Agent orchestration and LLM limitations
Single- and multi-agent applications, tool calling and LLM limitations are explicit requirements. Revision exercise: compare a single-agent and multi-agent design for one business workflow, including when an agent should call a tool and when it should stop.
Evaluation, guardrails and responsible AI
The role calls for evaluating response quality, reducing hallucinations and applying guardrails. Revision exercise: define a small set of test cases and explain what each would reveal about answer quality, unsupported claims and tool misuse.
Data and AI foundations
The posting requires prior experience in at least one of data engineering, the data science or ML lifecycle, or analytics engineering; it does not require all three. Revision exercise: trace how changes to an ETL/ELT pipeline, NLP workflow or data product could affect the data supplied to an LLM application.
Production integration and operations
The posting names enterprise integration, deployment, scaling, monitoring, containers and CI/CD. Revision exercise: sketch the path from a data platform through an LLM service to an enterprise workflow, and identify what you would monitor at each boundary.
Enterprise GenAI privacy
Data masking, access control and compliance are preferred, not must-have qualifications. Suggested revision exercise: identify where sensitive data might enter a RAG flow and describe where masking or access checks could apply.

Skills to revise

Build and tune a RAG pipeline
Hands-on RAG and retrieval optimisation are required. Practice ingesting a small document set, generating embeddings, retrieving passages and synthesising answers; compare retrieval results for several queries and explain one adjustment.
Implement an agent workflow
The role names LangChain, LangGraph or similar frameworks, plus prompt chaining, tool/function calling and structured outputs. Practice a short workflow with a tool call and a structured response; test how it handles an invalid tool result.
Develop a Python GenAI service
Production-grade Python, API integration and GenAI services are explicit requirements. Practice exposing a small LLM workflow through FastAPI or Flask, including input validation, integration-error handling and a reusable component.
Process and integrate large-volume data
The posting calls for PySpark, large-volume data analysis and integration with Fabric, Azure Databricks or Snowflake. Practice a PySpark transformation with data-quality checks, then explain how its output would feed or refresh an LLM application.
Run response evaluations
Evaluation frameworks and guardrails are part of the responsibilities. Practice running a fixed set of prompts through an application, recording retrieval relevance and unsupported-answer failures, then comparing results after one change.
Prepare a deployment and monitoring exercise
The posting expects exposure to cloud platforms, containers, CI/CD and monitoring, and collaboration with MLOps. As preparation, outline a containerised service release and a monitoring plan for API failures, response quality and scaling; these are exercise choices, not a stated employer process.

Your preparation checklist

Map evidence to the core requirements
The posting asks for 6–9 years of total experience, including 1–3+ years of production GenAI work, plus a foundation in data engineering, ML/NLP, or analytics engineering. Prepare a brief timeline and specific examples that substantiate whichever requirements apply.
Prepare a production RAG walkthrough
The role covers ingestion through response synthesis and retrieval optimisation. Prepare a diagram of a project you can discuss, showing data flow, embedding and retrieval choices, integration points, and measured improvements; distinguish implemented work from proposed changes.
Practice an agent-system design exercise
Sketch a single- or multi-agent business workflow using prompt chaining, tool or function calling, and structured outputs. Be ready to explain orchestration choices, failure handling, and when a simpler non-agent approach would suffice.
Show the data and API integration path
The posting calls for production Python/PySpark, API integration, large-volume data work, and integration with Fabric, Azure Databricks, or Snowflake. Prepare a concrete example showing data transformations, service interfaces, and your own contribution; use a hypothetical design where direct experience is unavailable.
Prepare an evaluation and operations plan
Draft a sample test set and measures for retrieval and answer quality, hallucinations, and guardrail behaviour. Explain how you would work with Data Engineering and MLOps on deployment, monitoring, scaling, documentation, and reusable components; label proposed practices as proposals.
Frame preferred topics accurately
If relevant, prepare a short account of actual work with LoRA/PEFT or prompt tuning, data masking and access control, Azure AI services, or agentic coding tools. These are preferred in the posting, so avoid presenting them as mandatory or claiming experience you cannot demonstrate.
Write targeted questions for the employer
Prepare questions about the business use cases, available data and enterprise integrations, how RAG and agent quality are evaluated, and the division of responsibilities with Data Engineering and MLOps. These clarify matters the posting does not specify.

Availability last checked: 4 Nov 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in India. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…