New opportunity

AI Scientist

Epergne Solutions · Bengaluru, India

About this role

The AI Scientist will conduct advanced research and development in generative AI, large language models, multimodal AI and computer vision. The role is based in Bengaluru, India, and calls for 7+ years of experience. The scientist will turn research into prototypes and production-ready models, improve large-scale models and datasets, and work with engineering and product teams on scalable AI systems.

Responsibilities

Implement and customise AI research from leading publications; experiment with fine-tuning, prompt engineering, architecture optimisation and evaluation to improve model performance.
Develop training and evaluation datasets through synthetic data generation, annotation workflows and benchmark creation.
Integrate research into production systems, contribute to research planning, and keep current with advances in the field.
Document experiments and technical findings; contribute to patents, technical reports and reusable research assets while maintaining reproducibility, benchmarking and responsible AI practices.

Requirements

A Master’s or PhD in Computer Science, Artificial Intelligence, Machine Learning or a related field, with experience in AI research, machine learning, deep learning or advanced AI development and demonstrated work on prototypes, research models or production-ready systems.
Strong expertise in LLMs, generative AI, multimodal AI and computer vision; proficiency in Python and frameworks such as PyTorch, TensorFlow or JAX.
Experience with tools such as Hugging Face Transformers, Diffusers, MLflow or TensorBoard; knowledge of distributed training frameworks such as DeepSpeed, FSDP or Horovod; and model optimisation techniques including quantisation, pruning and knowledge distillation.
Familiarity with data tools such as DVC, Apache Arrow or Label Studio, plus evaluation metrics, experimentation, benchmarking and reproducible research.

Preferred

Publications, patents or AI research or open-source contributions are advantageous. The posting also seeks analytical problem-solving, a research-oriented approach, cross-functional collaboration and clear technical communication and documentation.

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

Skills for this role

AI researchGenerative AILarge Language ModelsMultimodal AIComputer VisionDeep LearningPythonPyTorchTensorFlowJAXHugging Face TransformersDiffusersMLflowTensorBoardDeepSpeedFSDPHorovodDVCApache ArrowLabel StudioFine-tuningPrompt EngineeringDistributed TrainingQuantisationPruningKnowledge DistillationSynthetic Data GenerationData AnnotationBenchmarkingModel Evaluation Metrics

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

Turn research into working models
Implement and customise published AI research as prototypes and production-ready models across LLMs, generative AI, multimodal AI and computer vision.
Improve model performance through experimentation
Run and evaluate experiments using fine-tuning, prompt engineering and architecture optimisation; the posting also calls for large-scale model optimisation.
Build trustworthy datasets and benchmarks
Develop training and evaluation datasets through synthetic data generation, annotation workflows and benchmark creation, with reproducibility and responsible AI practices.
Integrate and communicate research
Work with engineering and product teams on scalable production systems; document experiments and findings, and contribute to research planning and reusable assets.

Likely challenges to think through

Research gains versus production constraints
Likely preparation challenge, not a stated employer problem: a promising result may be difficult to scale or serve efficiently. Prepare to explain how you would compare model quality with compute and deployment costs, including when quantisation, pruning or distillation might help.
Benchmark gains versus reliable evidence
Likely preparation challenge: fine-tuning or synthetic data could improve a benchmark without improving broader performance. Prepare to explain data-quality checks, evaluation design and reproducible experiments that would test whether a gain holds up.
Novelty versus implementation time
Likely preparation challenge: keeping pace with new research while delivering usable models requires selecting what to pursue. Prepare to describe a decision framework for reproducing a paper, measuring its value and communicating a recommendation to engineering and product teams.

Practice your answers

Technical practice

How would you turn a published AI method into a reproducible prototype?

What to coverThe role includes implementing and customising research papers. Suggested preparation: explain how you would identify essential design choices, establish a baseline, verify the reported result and document any deviations.

How would you decide between prompt engineering and fine-tuning for an LLM task?

What to coverThe posting names both approaches as ways to improve model performance. Suggested preparation: compare their data, compute and maintenance costs, then describe an experiment that would test the choice.

How would you design and evaluate a multimodal or computer-vision model?

What to coverMultimodal AI and computer vision are stated areas of expertise. Suggested preparation: explain how you would align inputs and outputs, select a suitable model approach and check performance across relevant input types without assuming a specific application.

When would you use a diffusion-based approach for generative AI?

What to coverThe posting covers generative AI and names Diffusers among example tools. Suggested preparation: discuss the output being generated, how you would assess quality and the trade-offs against another plausible approach.

How would you troubleshoot a large-model distributed training run that scales poorly?

What to coverThe posting calls for knowledge of frameworks such as DeepSpeed, FSDP or Horovod. Suggested preparation: describe how you would isolate memory, communication and data-loading bottlenecks, then measure whether a change improves throughput without altering model quality.

How would you choose among quantisation, pruning and knowledge distillation?

What to coverAll three model-optimisation techniques appear in the required skills. Suggested preparation: compare their effects on model size, runtime and quality, and explain how you would validate the resulting model against a baseline.

How would you build trustworthy training and evaluation datasets?

What to coverThe role includes synthetic data generation, annotation workflows and benchmark creation. Suggested preparation: address data quality checks, annotation consistency, leakage prevention and versioning; DVC, Apache Arrow and Label Studio are examples named in the posting.

How would you construct a benchmark for a model improvement?

What to coverBenchmarking, evaluation metrics and responsible AI practices are explicit responsibilities. Suggested preparation: define a baseline, choose metrics tied to the task, inspect failure cases and check whether aggregate gains conceal important regressions.

How would you make a series of model experiments reproducible?

What to coverThe posting calls for experiment documentation and reproducible research, and names MLflow and TensorBoard among example tools. Suggested preparation: cover configuration, data and model versions, seeds, recorded metrics and the information another researcher would need to rerun a comparison.

What would you change when moving a research prototype toward production?

What to coverThe role involves production-ready models and collaboration with engineering and product teams on scalable systems. Suggested preparation: explain how you would test reliability and performance, document limitations and agree on acceptance criteria before integration.

Behavioural practice

Tell me about a time you worked across research, engineering and product teams.

What to coverCross-functional integration is part of the advertised role. Suggested preparation: use a real situation/action/result example that shows how you reconciled differing priorities and what the collaboration produced.

Describe a research experiment that did not support your initial hypothesis.

What to coverThe posting emphasises analytical problem-solving and experimentation. Suggested preparation: give the situation, the tests and decisions you made, and the result or change in research direction.

Tell me about a time you improved the reproducibility of research work.

What to coverReproducibility and documentation are stated responsibilities. Suggested preparation: describe the initial problem, the concrete practice you introduced and evidence of its effect on subsequent work.

Describe how you communicated a complex technical finding to another audience.

What to coverThe role calls for technical reports, reusable research assets and clear communication. Suggested preparation: identify the audience and situation, explain how you adapted the message, and state the resulting decision or use of the finding.

Tell me about a time you helped prioritise competing research directions.

What to coverThe posting includes contributing to research planning and staying current with advances. Suggested preparation: give a real situation/action/result example showing the evidence and trade-offs you used, the choice made and what followed.

Sharpen your knowledge

Topics to prepare

LLM and generative AI methods
Review how fine-tuning, prompt engineering and architecture changes can affect model performance. As preparation, compare when you would test each approach; the posting explicitly calls for these experiments.
Multimodal AI and computer vision
Sketch how a model would combine image and text inputs, then identify likely evaluation challenges. These are named areas of expertise in the posting; the example task is a preparation suggestion, not a stated project.
Research-to-prototype decisions
Choose a published AI method and outline its key assumptions, implementation decisions and tests needed before production use. The role calls for adapting research papers into prototypes and production-ready models.
Scaling and model optimisation
Compare distributed training with quantisation, pruning and knowledge distillation: explain what each changes and which performance trade-offs you would measure. The posting names distributed frameworks and all three optimisation techniques.
Dataset design
Design a training and evaluation data plan covering synthetic examples, annotation checks and benchmark creation. The posting explicitly assigns development and management of these datasets.
Evaluation and responsible research
Draft a comparison plan with appropriate metrics, repeatable runs and checks for responsible AI concerns. The posting emphasises benchmarking, reproducibility and responsible AI practices.

Skills to revise

Implement a paper in Python
As practice, reproduce a small published model component using Python and one listed framework—PyTorch, TensorFlow or JAX—and record deviations from the paper. This mirrors the required research-implementation work.
Run a controlled model-improvement experiment
Test one fine-tuning or prompt-engineering change against a baseline, keeping the evaluation set fixed and recording the result. Both approaches and performance evaluation appear in the responsibilities.
Build a traceable dataset workflow
Prepare a small annotated or synthetic dataset, document its construction and create a separate benchmark split. If useful, practise with a listed tool such as DVC, Apache Arrow or Label Studio; the posting asks for familiarity, not every tool.
Exercise distributed training and optimisation
As a preparation exercise, run or outline a small distributed-training test using a listed framework such as FSDP, then compare an optimised model with its baseline. The role specifies distributed-framework knowledge and model optimisation experience.
Track and communicate experimental results
Log configurations, metrics and findings with a listed tool such as MLflow or TensorBoard, then write a concise technical report that another researcher could reproduce. The posting calls for experimentation, documentation and reproducibility.
Prepare a research-to-production handoff
Draft a handoff for a prototype covering its benchmark results, limitations and integration considerations. This is suggested practice for the stated collaboration with engineering and product teams on scalable production systems.

Your preparation checklist

Map the stated requirements to evidence
Make a one-page inventory of relevant research and development work, Python and framework experience, education, and experience length. The posting specifies 7+ years and a related Master’s or PhD; identify evidence to discuss without overstating it.
Prepare a research-to-model walkthrough
Select a relevant LLM, generative AI, multimodal, or computer-vision example, if available. Explain the paper or approach, your implementation choices, experiments, results, and what changed between prototype and production-ready model.
Build an evaluation and reproducibility brief
Prepare a concise account of a fine-tuning, prompting, or architecture experiment: baseline, metrics, benchmark, controls, tracked configurations, findings, and limitations. The role explicitly emphasizes evaluation and reproducible research.
Sketch a dataset and benchmark plan
For a plausible model task, outline training and evaluation data, synthetic-data use, annotation checks, versioning, and leakage risks. Treat the task as interview practice, not as a known company project.
Practice a scalable-model design exercise
Draw a high-level route from experiment to deployment, covering distributed training, an optimisation trade-off such as quantisation versus accuracy, and how engineering and product teams would validate the result.
Gather communication and research artifacts
Choose a technical report, experiment record, reusable asset, or—if applicable—publication, patent, or open-source contribution. Be ready to explain your contribution, decisions, responsible-AI considerations, and how findings were shared.
Write focused questions for the employer
Ask which model and data problems are priorities, how research is evaluated for production integration, what compute and benchmarking constraints apply, and how research, engineering, and product teams collaborate.

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…