New opportunity

Data Science Practitioner

Accenture · Hyderabad, India

About this role

Accenture seeks a Data Science Practitioner in Hyderabad to design and deliver AI/ML decision-making frameworks and demonstrate their business value. The detailed role focuses on building and productionizing NLP and generative AI capabilities for real-time, high-scale enterprise applications, including LLM-based agents, retrieval-augmented generation (RAG) and conversational AI.

Responsibilities

Design, fine-tune and optimize prompts for summarization, classification, sentiment analysis, Q&A, translation and knowledge-gap analysis. Build agentic workflows and RAG pipelines for multi-step evaluation and validation.
Apply classical NLP methods and transformer models alongside LLMs. Develop production inference APIs, integrate speech-to-text and text-to-speech capabilities, and select open-source or hosted models according to cost, latency and accuracy.
Monitor productivity and quality metrics, improve models through tuning, feedback and production-data evaluation, and present KPIs and insights to business stakeholders through dashboards. Work with data engineering on ingestion and synchronization across relational and vector stores.

Requirements

The posting lists a minimum of 5 years of experience in its project requirements and separately specifies 6–7 years, including 6+ years in NLP, data science or AI engineering with production deployment experience, for the detailed role.
Strong Python and experience serving ML/AI models through production APIs; NLP fundamentals, transformer models, supervised and unsupervised learning, and evaluation methods. Experience with prompt engineering, RAG, agentic frameworks and LLM APIs; SQL databases and exposure to a vector store. Hands-on experience with Azure AI Foundry, Azure Function Apps and Azure Logic Apps, or equivalent AWS or GCP AI and serverless services.
The project requirements specify 15 years of full-time education. The detailed role lists a bachelor’s or master’s degree in data science, computer science, statistics or a related field.

Preferred

Knowledge graphs or graph-based retrieval such as Neo4j; production deployments across AWS, Azure and GCP; speech integration; low-code or no-code agent prototyping with tools such as Microsoft Power Platform or Copilot Studio; Power BI or Tableau; and exposure to healthcare or other regulated-data domains.

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

Skills for this role

Data SciencePythonNatural Language ProcessingGenerative AILarge Language ModelsPrompt EngineeringRetrieval-Augmented GenerationConversational AIAI AgentsCrewAILangChainLangGraphText ClassificationNamed Entity RecognitionClusteringTopic ModelingSummarizationSentiment AnalysisTransformersBERTRoBERTaT5BARTGPTClaudeSupervised LearningUnsupervised LearningModel EvaluationHyperparameter TuningProduction APIs for ML Models（修正）

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

Deliver production NLP and generative AI features
The role calls for prompts for summarization, classification, sentiment, Q&A and other tasks, plus RAG and agentic workflows for multi-step evaluation. Classical NLP and transformer models are also part of the stated work.
Serve AI capabilities at enterprise scale
The posting emphasizes real-time inference APIs for high-scale applications, production deployment, and hands-on Azure AI and serverless services or AWS/GCP equivalents. Speech integration is among the listed responsibilities.
Measure performance and business value
The project role requires measuring and justifying AI/ML solution value. The detailed role calls for monitoring productivity and quality, evaluating against production data, iterating through tuning and feedback, and presenting KPIs to business stakeholders.
Make model and data architecture choices
The role requires selecting open-source or hosted LLMs against cost, latency and accuracy needs, and partnering with data engineering on ingestion and synchronization across relational and vector stores.

Likely challenges to think through

Balancing response quality with speed and cost
Preparation hypothesis, not a reported employer problem: real-time, high-scale APIs may make a more capable model or multi-step agent workflow too slow or costly. Prepare to explain how you would compare models and workflow designs using accuracy, latency and cost measurements.
Keeping retrieval useful as data changes
Preparation hypothesis: synchronization between relational and vector stores may complicate RAG freshness and validation. Prepare to describe how you would check retrieval quality, detect stale results and evaluate the effect on answers.
Connecting model metrics to business outcomes
Preparation hypothesis: better AI output scores may not automatically demonstrate productivity or business value. Prepare to distinguish output-quality measures from operational and business KPIs, and explain how production feedback would guide iteration.

Practice your answers

Technical practice

How would you define and measure the business value of an AI/ML decision-making framework?

What to coverThe role explicitly calls for delivering decision-making frameworks and justifying their value. Suggested preparation: explain how you would establish a baseline, choose outcome and quality measures, and account for the costs of incorrect decisions.

How would you design and evaluate prompts for summarization, classification, Q&A, and translation?

What to coverThese are among the production LLM features named in the posting. Suggested preparation: discuss task-specific instructions, representative test cases, failure analysis, and when prompt changes are insufficient.

When would you use classical NLP, a transformer model, or an LLM for a text task?

What to coverThe role calls for classical NLP techniques and transformer models alongside LLMs. Suggested preparation: compare approaches for tasks such as classification, named entity recognition, and topic modeling using accuracy, data needs, latency, and maintainability.

How would you build and troubleshoot a RAG pipeline using relational and vector stores?

What to coverThe posting describes RAG and ingestion or synchronization across relational and vector stores. Suggested preparation: trace document updates through indexing and retrieval, then explain how you would investigate stale results or irrelevant retrieved context.

How would you validate a multi-step agentic workflow?

What to coverThe role includes agent orchestration and multi-step evaluation and validation. Suggested preparation: explain how you would define each step's inputs and outputs, test intermediate results, and handle failures before accepting a final answer.

How would you choose between an open-source and a hosted LLM?

What to coverThe posting explicitly names cost, latency, and accuracy as selection criteria. Suggested preparation: describe a comparative evaluation on representative workloads, including how deployment and operational constraints could change the choice.

How would you serve an NLP or generative AI feature through a real-time production API?

What to coverStrong Python API experience, scaled inference, and Azure AI and compute services or equivalent AWS/GCP services are listed. Suggested preparation: cover request handling, latency and capacity limits, failure behavior, and what you would monitor in a serverless or comparable deployment.

How would you detect and address declining model quality after deployment?

What to coverThe role calls for monitoring quality and productivity, using feedback loops, and evaluating against production data. Suggested preparation: distinguish changes in input data from model failures, explain how you would review feedback, and describe when tuning or another intervention is warranted.

What would you test when adding speech-to-text and text-to-speech to conversational AI?

What to coverSpeech integration is a stated responsibility. Suggested preparation: consider transcription errors, response quality, end-to-end latency, and how errors in one stage affect the conversation.

When might graph-based retrieval improve a RAG system?

What to coverKnowledge graphs and graph-based retrieval are listed as preferred, not required. Suggested preparation: contrast relationship-based queries with vector retrieval and explain how you would test whether the added complexity improves results.

Behavioural practice

Tell me about a time you coordinated with data engineering on an AI data pipeline.

What to coverThe role involves partnering on ingestion and synchronization. As a suggested practice answer, describe the situation, your actions to resolve a data or handoff issue, and the resulting effect on delivery or data quality.

Describe a time you explained model performance to business stakeholders.

What to coverThe posting calls for presenting KPIs and insights through dashboards. As a suggested practice answer, give the situation, how you chose and communicated the measures, and what decision or understanding resulted.

Tell me about a production AI quality or reliability problem you helped resolve.

What to coverProduction deployment and ongoing performance improvement are central to the role. As a suggested practice answer, set out the situation, your investigation and actions, and the measured result or lesson learned.

Describe a time you made a trade-off between model quality, latency, and cost.

What to coverThe posting explicitly identifies those factors in model and deployment selection. As a suggested practice answer, explain the situation, the evidence behind your choice, and its outcome.

Tell me about a time feedback changed an AI feature after its initial release.

What to coverThe role calls for feedback loops and evaluation against production data. As a suggested practice answer, describe the situation, how you assessed and acted on the feedback, and what changed as a result.

Sharpen your knowledge

Topics to prepare

NLP methods and transformer selection
Suggested revision exercise: Explain when you would use text classification, NER, clustering or topic modeling, then compare a transformer approach with an LLM approach for one task. The posting calls for classical NLP techniques and transformer models alongside LLMs.
RAG and agentic workflow design
Suggested revision exercise: Diagram a question-answering flow from ingestion and retrieval through an agent’s multi-step validation to a final answer; identify where a relational store and vector store fit. The role explicitly covers RAG, agentic workflows and both store types.
Evaluation and improvement
Suggested revision exercise: Define a small evaluation set and explain how you would assess classification quality and answer quality, then use feedback and production data to decide what to tune. The posting requires evaluation methodology and performance iteration.
Model and deployment trade-offs
Suggested revision exercise: Compare an open-source and a hosted LLM for a real-time use case, stating how cost, latency and accuracy would affect your choice. Selecting models and deployments on those grounds is an explicit responsibility.
Business value and AI-output metrics
Suggested revision exercise: Choose one productivity KPI and one quality KPI for a conversational AI feature, and explain how each supports a business-outcome discussion. The role calls for measuring solution value and presenting KPIs to stakeholders.

Skills to revise

Serve a model through a Python API
Suggested practice: Build a small Python inference endpoint for text classification or Q&A, and test response time and failure handling. Strong Python and production inference APIs are explicit requirements.
Test task-specific prompts
Suggested practice: Write two prompt variants for summarization or sentiment analysis, run both on the same sample texts, and record output-quality differences. Prompt design and optimization for these features are stated responsibilities.
Build a small RAG workflow
Suggested practice: Index a small document set in a vector store, retrieve relevant passages, and use an LLM API to answer questions with a validation step. This exercises the posting’s RAG, vector-store and multi-step workflow requirements.
Reconcile relational and vector data
Suggested practice: Map document IDs and update timestamps between a SQL table and vector-store records; demonstrate how an updated document would be synchronized. The role involves partnering on ingestion and synchronization across these stores.
Prototype cloud-backed inference
Suggested practice: If you have access, prototype an inference flow using Azure AI Foundry and Azure Function Apps, noting where Azure Logic Apps would fit; an equivalent AWS or GCP AI and serverless stack is also listed. The posting requires hands-on depth with one of these options.
Report model performance
Suggested practice: Create a simple view of evaluation results, latency and a productivity metric, then rehearse a short explanation of the trade-offs for a business stakeholder. Monitoring AI-output metrics and presenting performance through dashboards are stated responsibilities.

Your preparation checklist

Map the stated requirements to evidence
Prepare a one-page list of examples you can substantiate for Python, production AI APIs, NLP, transformers, RAG, agentic tooling, SQL, a vector store and Azure services or AWS/GCP equivalents. Keep the posting’s minimum five years and detailed six-plus-years statements distinct.
Rehearse a production project walkthrough
If you have a relevant project, prepare a concise account of the problem, your decisions, deployment, evaluation, trade-offs and business result. Distinguish your own contributions from the team’s; label any alternative design as hypothetical.
Draw a RAG and agent workflow
As an interview exercise, diagram ingestion and synchronization across relational and vector stores, retrieval, multi-step agent orchestration, validation and a production inference API. Explain how you would handle real-time scale and conversational use cases.
Prepare an NLP model comparison
Use a sample task from the posting, such as classification or summarization, to explain when you would choose classical methods, a transformer or an LLM. Bring an evaluation plan covering prompt changes, supervised or unsupervised methods where relevant, and error analysis.
Work through deployment trade-offs
Sketch how you would serve the design using Azure AI and compute services or equivalent AWS/GCP services. State what you would measure to choose an open-source or hosted model for cost, latency and accuracy; include speech integration only if relevant to your example.
Show how you would report value
Prepare a sample KPI dashboard outline linking output quality and productivity measures to business outcomes. For any real results you discuss, use documented figures and explain how production feedback would guide improvements.
Bring targeted questions for the employer
Ask which experience and education statements govern this role, what scale and latency the applications require, which cloud and data stores are in use, and how stakeholders measure AI solution value.

Availability last checked: 5 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…