New opportunity

Artificial Intelligence Researcher

Antal International · Bangalore, India

About this role

This permanent, on-site role in Bangalore focuses on research into agentic reasoning and decision intelligence, connecting scientific advances with practical AI applications. The position is posted by Antal International for a client.

Responsibilities

Research ways to improve the reliability and learning of agentic reasoning systems, and define agent–environment–reward formulations for decision intelligence workflows.
Frame learning problems using deep reinforcement learning, preference learning or supervised fine-tuning. Design and evaluate Chain-of-Thought, Tree-of-Thought and multi-step planning approaches.
Curate datasets to train and evaluate reasoning agents. Contribute to knowledge-system learning through graph updates and ontology refinement.
Work with AI engineers to translate research results into production systems.

Requirements

A bachelor’s or master’s degree in computer science, AI, data science or a related field. An advanced degree or strong academic research background is preferred.
The listing states 6–10 years of experience; its detailed requirements specify 2–6+ years in data science, applied research or ML engineering. Hands-on experience with deep learning frameworks and reinforcement learning projects is required.
Strong PyTorch and TensorFlow proficiency; working knowledge of reinforcement learning algorithms such as MDPs, PPO, DPO and GRPO; and experience with transformers and LLM fine-tuning, including SFT, LoRA and QLoRA.
Understanding of classical machine learning and statistical learning, strong analytical and experimental-design skills, and the ability to evaluate AI agents rigorously. The role calls for scientific curiosity, collaboration and comfort with open-ended problems while pursuing practical results.

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

Skills for this role

AI researchData ScienceMachine LearningPythonLarge Language ModelsTensorFlowDeep LearningPyTorchReinforcement LearningAgentic AIDecision intelligencePreference learningSupervised fine-tuningChain-of-ThoughtTree-of-ThoughtMulti-step planningDataset curationKnowledge graphsOntology refinementMarkov decision processesPPODPOGRPOTransformersLoRAQLoRAStatistical learningExperimental designAI agent evaluation

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

Reliable agentic reasoning
The role calls for research that improves agent reliability and learning. Prepare to explain how you would define an agent, its environment and its rewards for a decision-intelligence workflow, then test whether the formulation improves behavior.
Learning and planning experiments
The posting names deep reinforcement learning, preference learning and supervised fine-tuning, alongside Chain-of-Thought, Tree-of-Thought and multi-step planning. Prepare to compare when you would use each approach and how you would evaluate it rigorously.
Datasets and knowledge systems
Curating data for agent training and evaluation, plus graph updates and ontology refinement, are stated duties. Prepare to discuss data quality, evaluation-set design and how changes to a knowledge system could be validated.
Research with practical impact
The researcher will work with AI engineers to translate findings into production systems. Prepare an example of communicating experimental results, limitations and implementation considerations to engineers.

Likely challenges to think through

Reward design versus intended behavior
Preparation hypothesis: A measurable reward may not capture the decision a workflow actually needs. Practice explaining how you would detect misleading reward signals and assess agent reliability beyond a single score.
Choosing among reasoning approaches
Preparation hypothesis: Deeper planning or more elaborate training may improve some results while adding complexity. Prepare a controlled comparison of simpler and more involved approaches, with clear criteria for deciding whether an improvement is meaningful.
Moving from experiments to deployment
Preparation hypothesis: A promising research result may be difficult to reproduce or integrate. Prepare to explain what evidence, failure analysis and implementation details you would share with AI engineers before recommending adoption.

Practice your answers

Technical practice

How would you formulate an agent–environment–reward problem for a decision intelligence workflow?

What to coverSuggested preparation: Define the agent’s actions, observations, state assumptions and reward, then explain how the formulation captures the intended decision. Discuss how you would detect reward designs that encourage unreliable behaviour.

When would you frame an agent-learning problem as deep reinforcement learning, preference learning or supervised fine-tuning?

What to coverSuggested preparation: Compare the feedback and training data each approach needs and what each can optimise. Explain how you would choose among them for a stated reasoning objective.

How do PPO, DPO and GRPO differ as approaches to improving an agent?

What to coverSuggested preparation: Explain their learning signals and the practical trade-offs relevant to reasoning agents. Clarify which assumptions and available data would affect your choice rather than treating the algorithms as interchangeable.

How would you compare Chain-of-Thought, Tree-of-Thought and multi-step planning?

What to coverSuggested preparation: Describe the reasoning structure each provides and where additional search or planning complexity might be justified. Propose a comparison that measures answer quality and reliability alongside computational cost.

How would you curate a dataset for training and evaluating reasoning agents?

What to coverSuggested preparation: Address task coverage, data quality, leakage and separation of training from evaluation examples. Explain how the dataset would expose failures in multi-step reasoning rather than only measure easy cases.

How would you design a rigorous experiment to test whether an agent has become more reliable?

What to coverSuggested preparation: Specify baselines, meaningful metrics, controlled comparisons and analysis of variation or uncertainty. Include failure-case analysis and explain what evidence would support a real improvement rather than a noisy result.

How might graph updates and ontology refinement contribute to a learning knowledge system?

What to coverSuggested preparation: Explain how changes to relationships or concepts could affect an agent’s reasoning and how you would check their consistency. Discuss a way to evaluate whether an update helps the intended tasks without introducing new errors.

How would you choose between full supervised fine-tuning, LoRA and QLoRA for a transformer?

What to coverSuggested preparation: Compare adaptation scope, resource demands and potential quality trade-offs for an LLM reasoning task. Explain how you would evaluate the fine-tuned model against the original objective.

How would you structure a reproducible agent-learning experiment in PyTorch or TensorFlow?

What to coverSuggested preparation: Cover implementation checks, controlled runs and tracking of training and evaluation results. Explain how you would troubleshoot a result that changes unexpectedly between runs or frameworks.

What would you validate before translating a reasoning research result into a production system?

What to coverSuggested preparation: Identify the assumptions, evaluation evidence and failure modes AI engineers would need to understand. Discuss the trade-off between the research result’s performance and its operational complexity without assuming a specific production stack.

Behavioural practice

Tell me about a time you turned an open-ended research question into a tractable experiment.

What to coverSuggested preparation: Give the situation, the hypothesis and actions you took to narrow the question, and the result. Explain what the evidence changed about your next step.

Describe a time an experiment contradicted your expectations.

What to coverSuggested preparation: Set out the situation, how you checked the data or methodology, and what you changed. State the result and what you learned without presenting an inconclusive finding as a success.

Tell me about a time you worked with engineers to make a research result practically usable.

What to coverSuggested preparation: Describe the situation, your contribution and how you communicated limitations or trade-offs. Give the result, including what remained unresolved if applicable.

Describe a time you had to balance scientific exploration with practical impact.

What to coverSuggested preparation: Explain the competing priorities, the action you chose and your reasoning. Describe the result and how you determined whether the work was useful.

Sharpen your knowledge

Topics to prepare

Agent–environment–reward formulation
Posting: The role defines these formulations for decision intelligence workflows and calls for knowledge of MDPs. Suggested exercise: Describe an example workflow in terms of states, actions, transitions and rewards, then identify where the reward might produce unintended behavior.
Learning objectives and algorithms
Posting: The role frames problems using deep reinforcement learning, preference learning or supervised fine-tuning, and names PPO, DPO and GRPO. Suggested exercise: Compare the training signal, data needed and appropriate use of each approach for a reasoning agent.
Reasoning and planning paradigms
Posting: Researchers design and evaluate Chain-of-Thought, Tree-of-Thought and multi-step planning approaches. Suggested exercise: Work through one decision problem using each approach and specify how you would compare reliability.
Transformers and LLM adaptation
Posting: Experience with transformers and SFT, LoRA and QLoRA is requested. Suggested exercise: Explain what each adaptation method changes, what data it needs and how you would check whether it improved reasoning.
Statistical learning and experimental design
Posting: The role requires classical and statistical learning concepts, experimental design and rigorous agent evaluation. Suggested exercise: Draft a comparison with a baseline, held-out cases, repeated trials and an analysis of failure types.
Knowledge-system learning
Posting: The work includes graph updates and ontology refinement. Suggested exercise: Sketch how a proposed graph update would be checked for consistency with existing entities and relationships.

Skills to revise

Implement models in both frameworks
Posting: Strong PyTorch and TensorFlow proficiency and hands-on deep learning experience are requested. Suggested exercise: Implement the same small training task in each framework and explain how you verified comparable results.
Prototype an agent learning setup
Posting: The role involves reinforcement learning projects and agent–environment–reward formulations. Suggested exercise: Build a small agent–environment loop, record outcomes and test whether changing the reward changes the intended behavior.
Curate a reasoning dataset
Posting: Researchers curate data for training and evaluating reasoning agents. Suggested exercise: Create a small set of reasoning cases, document inclusion rules and separate training from evaluation cases.
Run a controlled agent evaluation
Posting: The role calls for rigorous evaluation of reasoning agents. Suggested exercise: Compare two reasoning approaches on the same cases, record success and failure categories, and explain what the results do not establish.
Review graph and ontology changes
Posting: The work includes graph updates and ontology refinement. Suggested exercise: Make a sample update, check for duplicate or contradictory relationships, and document the rationale for accepting it.
Communicate a research handoff
Posting: Researchers collaborate with AI engineers to translate results into production systems. Suggested exercise: Write a brief handoff covering the tested approach, evidence, limitations and conditions that would warrant further evaluation.

Your preparation checklist

Prepare a research project walkthrough
Choose a relevant reinforcement learning or agentic-reasoning project. Bring a one-page account of the problem, your contribution, experimental design, results, limitations and next steps.
Draft an agent formulation
As a preparation exercise, define an agent, environment, actions and reward for a decision-intelligence workflow. Explain how you would test whether the formulation improves reliability.
Build a reasoning evaluation plan
Sketch an experiment comparing Chain-of-Thought, Tree-of-Thought and multi-step planning. Specify a baseline, evaluation data, reliability metrics and how you would inspect failures.
Gather technical evidence
Prepare concise examples of hands-on PyTorch and TensorFlow work, reinforcement learning projects, and transformer fine-tuning. Be ready to explain where MDPs, PPO, DPO, GRPO, SFT, LoRA or QLoRA fit—and which you have actually used.
Outline a data and knowledge-system approach
Describe how you would curate training and evaluation datasets for reasoning agents. Include checks for data quality and leakage, plus an approach to evaluating graph updates or ontology refinement.
Explain a research-to-production handoff
Prepare an example or proposed handoff that gives AI engineers reproducible results, known failure modes and clear criteria for deciding whether a research result is ready to use.
Prepare questions for Antal International
Ask what the client’s decision-intelligence workflows involve, how agent reliability is measured, and how researchers collaborate with AI engineers. Clarify the listing’s 6–10 years label versus its 2–6+ years detailed requirement.

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…