New opportunity

AI Research Engineer

Epergne Solutions · Bengaluru, India

About this role

The AI Research Engineer will design, develop and deploy scalable machine learning systems and AI-powered features in Bengaluru, India. The role spans LLM, computer vision and multimodal pipelines, with an emphasis on production reliability, performance, observability and cost efficiency.

Responsibilities

Build AI features from data ingestion through real-time inference, including training and inference pipelines for LLM, computer vision and multimodal models.
Develop scalable machine learning services and evaluation frameworks using offline evaluation, online experiments and user feedback.
Deploy, monitor and maintain models on containerised, cloud-based infrastructure. Investigate production incidents and improve reliability and operational performance.
Reduce training and inference costs through batching, quantization, mixed precision and GPU resource management. Collaborate with software engineering, data and product teams to deliver features.

Requirements

The posting states 7+ years of experience; its qualifications section separately lists 3–5+ years in applied machine learning, AI engineering or software engineering. A bachelor’s or master’s degree in Computer Science, Artificial Intelligence, Machine Learning or a related field is required.
Strong Python proficiency and experience with PyTorch or TensorFlow; end-to-end machine learning pipelines covering data preparation, training, evaluation, deployment and monitoring; and production-grade machine learning applications.
Knowledge of MLOps tools such as MLflow, Weights & Biases, DVC, Airflow or Prefect; Docker, Kubernetes, containerised deployment and CI/CD; and GPU optimization, ONNX, TensorRT, batching and mixed precision.
Familiarity with vector databases, retrieval-augmented generation and LLM fine-tuning; monitoring tools such as Prometheus, Grafana or OpenTelemetry; and statistics, experimentation and real-world model performance analysis. Analytical, problem-solving and collaboration skills are required.

Preferred attributes include hands-on LLM, computer vision or multimodal AI experience; balancing research innovation with production engineering; ownership in cross-functional teams; and strong communication and technical documentation skills.

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

Skills for this role

PythonPyTorchTensorFlowMachine learningMLOpsMLflowWeights & BiasesDVCAirflowPrefectDockerKubernetesCI/CDGPU optimizationONNXTensorRTBatchingQuantizationMixed precisionVector databasesRetrieval-augmented generationLLM fine-tuningPrometheusGrafanaOpenTelemetryModel evaluationExperimentationStatisticsComputer visionMultimodal AI systems

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 end-to-end AI features
The role covers data ingestion through real-time inference, including training and inference pipelines for LLM, computer vision, and multimodal models. Prepare to explain the architecture and production handoffs of an end-to-end system.
Measure model performance in use
The posting calls for evaluation frameworks combining offline evaluation, online experiments, and user feedback. Prepare to explain how these signals inform a deployment or iteration decision.
Own production reliability
Deploying, monitoring, and maintaining models on containerised cloud infrastructure—and investigating incidents—are explicit duties. Prepare an account of how you would detect, diagnose, and prevent a model-service failure.
Control compute cost and collaborate on delivery
The role calls for cost-efficient training and inference through batching, quantisation, mixed precision, and GPU resource management, alongside delivery with software engineering, data, and product teams. Prepare to explain both a measured optimisation decision and how its trade-offs would be communicated.

Likely challenges to think through

Balancing model quality against inference constraints
Likely preparation hypothesis, not a confirmed employer problem: real-time inference and cost optimisation can put quality, latency, and GPU spend in tension. Prepare a comparison that measures all three and explains when an optimisation should be rejected.
Reconciling offline and live results
Likely preparation hypothesis: offline evaluation may not reflect user feedback or online experiment results. Prepare to explain how you would investigate a mismatch, check experiment validity, and decide whether to revise or roll back a model.
Moving research work into reliable services
Likely preparation hypothesis: LLM, vision, or multimodal experimentation may compete with reproducibility and operational stability. Prepare to explain how you would move a promising model into a monitored, containerised service while coordinating changes across teams.

Practice your answers

Technical practice

How would you build an end-to-end training pipeline from data preparation to deployment?

What to coverSuggested preparation: Explain data validation, training in Python with PyTorch or TensorFlow, evaluation, and the handoff to deployment and monitoring. Discuss how you would make each stage repeatable and identify failures before release.

How would you design a scalable service for real-time model inference?

What to coverSuggested preparation: Describe the path from an incoming request through preprocessing, inference, and response. Explain how you would assess latency, throughput, reliability, and cost as demand changes.

When would you use retrieval-augmented generation rather than LLM fine-tuning?

What to coverSuggested preparation: Compare the approaches for incorporating task-specific information, including retrieval quality, update frequency, evaluation, and operational cost. Explain how a vector database fits into a retrieval pipeline.

How would you develop and evaluate a computer vision or multimodal pipeline?

What to coverSuggested preparation: Walk through data preparation, model training, inference, and evaluation for one relevant modality or combination of modalities. Explain how you would detect errors that an aggregate metric might conceal.

How would you combine offline evaluation, online experiments, and user feedback?

What to coverSuggested preparation: Describe what each source of evidence can reveal about model quality and real-world performance. Discuss experimental comparisons, statistical uncertainty, and how feedback would inform the next evaluation.

How would you make ML experiments and pipelines reproducible?

What to coverSuggested preparation: Explain how you would track datasets, code, configurations, and model results, and coordinate recurring pipeline steps. The posting names tools such as MLflow, Weights & Biases, DVC, Airflow, and Prefect as examples; focus on the responsibilities rather than claiming one is mandatory.

How would you deploy and update a model on containerised cloud infrastructure?

What to coverSuggested preparation: Cover packaging, CI/CD checks, deployment using Docker and Kubernetes, and verification after release. Discuss how you would reduce the risk of a model or service update disrupting inference.

How would you investigate a production inference incident?

What to coverSuggested preparation: Explain how metrics, logs, or traces would help distinguish service failures from model-performance problems. Discuss immediate mitigation, root-cause analysis, and a measurable reliability improvement.

How would you reduce GPU training and inference costs without losing necessary quality?

What to coverSuggested preparation: Compare batching, quantisation, mixed precision, and GPU resource management against memory use, throughput, latency, and model quality. Explain what you would measure before and after an optimisation.

When would you consider ONNX or TensorRT for inference?

What to coverSuggested preparation: Explain their potential role in improving deployment performance and how you would benchmark the result against the existing model. Discuss compatibility, numerical differences, and maintenance costs before adopting a conversion.

Behavioural practice

Tell me about a production ML incident you helped resolve.

What to coverSuggested preparation: Give the situation, your actions in investigation and coordination, and the result. Distinguish the immediate fix from any longer-term reliability improvement.

Describe a time you delivered an AI feature with software, data, and product colleagues.

What to coverSuggested preparation: Explain the situation, how you aligned requirements and responsibilities, and what you personally did. State the result and how you handled a cross-functional dependency or disagreement.

Tell me about a time you balanced a promising model improvement against production constraints.

What to coverSuggested preparation: Set out the situation, the quality, latency, reliability, or cost trade-off, and the evidence behind your decision. Explain your actions and the observed result without presenting a hypothetical outcome as experience.

Describe a time you communicated a model limitation or evaluation result to stakeholders.

What to coverSuggested preparation: Explain the situation, how you conveyed uncertainty and practical implications, and whether you documented the decision. Give the result, including any change to the feature or evaluation plan.

Sharpen your knowledge

Topics to prepare

End-to-end ML system design
The posting requires pipelines from data preparation through real-time inference, deployment and monitoring. Suggested exercise: sketch the stages of one production ML feature and explain how data and model versions move between them.
Evaluation and experimentation
The role calls for offline evaluation, online experiments, user feedback and real-world performance analysis. Suggested exercise: define offline metrics, an online experiment and a way to detect disagreement between the two.
Production reliability and observability
The posting emphasizes observable services, model monitoring and production incident investigation. Suggested exercise: trace a hypothetical inference-quality or latency incident from detection to diagnosis and prevention.
Inference and training efficiency
The role explicitly mentions batching, quantisation, mixed precision, GPU resource management, ONNX and TensorRT. Suggested exercise: compare when each technique might reduce cost and what accuracy, latency or operational trade-off to check.
LLM and multimodal pipeline choices
The posting covers LLM, computer vision and multimodal pipelines, with familiarity in vector databases, RAG and LLM fine-tuning. Suggested exercise: explain when retrieval or fine-tuning might be appropriate and how inputs and outputs would be evaluated; this is preparation, not a request to build every model type.

Skills to revise

Implement a model pipeline in Python
Strong Python and PyTorch or TensorFlow proficiency are required. Suggested exercise: implement data preparation, training, validation and inference for a small model, then explain how you would make its outputs deployable.
Track and reproduce experiments
The posting names MLflow, Weights & Biases, DVC, Airflow and Prefect as examples of MLOps tools. Suggested exercise: use one named tool to record a run or manage a pipeline, including the configuration and data version needed to reproduce a result.
Containerise and release an inference service
Docker, Kubernetes, containerised deployment and CI/CD are explicit requirements. Suggested exercise: package a small inference service with Docker and outline a CI/CD check and Kubernetes deployment and rollback plan.
Benchmark an optimisation change
The role calls for reducing training and inference costs. Suggested exercise: benchmark batching or mixed precision against a baseline, recording throughput, latency, resource use and any quality change.
Diagnose a monitored service
The posting names Prometheus, Grafana and OpenTelemetry as monitoring examples and expects incident investigation. Suggested exercise: choose one named tool, identify signals for an inference service and walk through a latency-spike diagnosis.
Prototype retrieval-augmented inference
Familiarity with vector databases and RAG is specified. Suggested exercise: assemble a small retrieval-to-answer flow and test cases where retrieval helps or supplies irrelevant context; a prototype is a preparation suggestion, not a stated job assignment.

Your preparation checklist

Map the stated requirements to evidence
Prepare a one-page matrix with examples you can substantiate for Python, PyTorch or TensorFlow, production ML pipelines, MLOps, containers and CI/CD. Note relevant degree and experience evidence without assuming you meet every requirement.
Rehearse an end-to-end project walkthrough
Choose a project you can discuss and diagram data preparation, training, evaluation, deployment and monitoring. Be ready to explain your decisions, measured results and individual contribution.
Practice a real-time AI system design
Sketch a hypothetical feature from data ingestion to inference using an LLM, computer vision or multimodal pipeline. Explain scaling, observability and, where relevant, retrieval or fine-tuning choices; this is a preparation exercise, not a stated project.
Draft an evaluation plan
Write a sample plan covering offline metrics, an online experiment, user feedback and analysis of real-world performance. State how the evidence would inform a deployment decision.
Prepare an operations and cost review
Outline how you would deploy and monitor a containerised cloud service, investigate an incident and compare cost or latency changes from batching, quantisation, mixed precision or GPU management.
Bring cross-functional examples and questions
Prepare a concise example of working with software, data or product teams, plus questions about feature priorities and production ownership. Ask which experience threshold applies: the posting says 7+ years, while its qualifications section says 3–5+ years.

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…