New opportunity

AI Research Internship

AryaXAI · Bangalore, India

About this role

This full-time, six-month AI research internship is described as a role with Lexsi Labs on the AryaXAI posting. The intern will collaborate with research and engineering teams on one frontier AI research area, with a rolling start date and potential extension or full-time offer. The posting lists Bangalore as the location and Remote as the working arrangement; applicant location requirements specify India.

Responsibilities

Depending on the selected research area, develop open-source Python tools for alignment, explainability, uncertainty quantification, robustness or machine unlearning; investigate model interpretability across text, images and tabular data; or probe internal model representations and failure modes through mechanistic interpretability.
Alternatively, develop and benchmark uncertainty and robustness methods for foundation models; improve tabular foundation model architectures and the TabTune library; or explore reinforcement-learning algorithms and an RL fine-tuning library.
Maintain experimental code, run systematic studies and contribute to whitepapers or conference submissions.

Required qualifications

Strong Python skills and the ability to write clean, modular, testable code; a deep understanding of machine learning and deep learning principles and hands-on PyTorch experience.
Comprehensive knowledge of transformer fundamentals, including attention, positional encodings, tokenization and training objectives across architectures such as BERT, GPT, LLaMA, T5, MoE and Mamba.
Familiarity with Git workflows, CI/CD, packaging and documentation, alongside communication, code-review and collaborative teamwork skills.

Preferred qualifications

Expertise in at least one listed domain: explainability; mechanistic interpretability; uncertainty estimation; quantization and pruning; LLM alignment, prompting and RLHF; tabular foundation models; or fine-tuning and post-training adaptation.
Additional advantages include research publications or equivalent experience, open-source AI/ML contributions, exposure to risk-sensitive applications such as finance or healthcare, and familiarity with large-scale training infrastructure.

The posting offers a competitive stipend without a stated amount, access to GPUs, cloud credits and proprietary models, and opportunities for research co-authorship.

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

Skills for this role

PythonPyTorchMachine LearningDeep LearningTransformer ArchitecturesGitCI/CDExplainable AISHAPLIMEIntegrated GradientsLRPGrad-CAMMechanistic InterpretabilityActivation PatchingFeature VisualizationUncertainty QuantificationBayesian MethodsEnsemble MethodsTest-Time AugmentationReinforcement LearningRLHFReward ModelingModel AlignmentModel RobustnessMachine UnlearningTabular Foundation ModelsTabTuneTabPFNTabICL Island?

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

Make progress in one research area
Posting: The intern works with research and engineering teams on one area, such as interpretability, uncertainty and robustness, tabular foundation models, or RL. Prepare: Choose one preferred area and explain a specific research direction within it; the posting does not expect work across every track.
Design systematic studies
Posting: Duties include running systematic studies and maintaining experiment code. Prepare: Be ready to explain how you would select baselines, metrics and controls, and interpret results that do not support a hypothesis.
Build reusable research software
Posting: Possible work includes open-source Python tooling and libraries such as TabTune or an RL fine-tuning library. Clean, modular, testable code, Git, packaging and documentation are required. Prepare: Discuss how you would turn an experiment into code others can review and use.
Apply strong model foundations
Posting: The role requires ML and deep-learning principles, hands-on PyTorch, and transformer fundamentals including attention, positional encodings, tokenization and training objectives. Prepare: Connect these fundamentals to the methods in your preferred research area rather than reciting architecture names.
Communicate findings collaboratively
Posting: The intern collaborates with research and engineering teams, participates in peer code review, and may co-author whitepapers or conference submissions. Prepare: Practice presenting a method, its evidence and its limitations clearly to both audiences.

Likely challenges to think through

Narrow a broad frontier topic
Likely challenge (inference): Each listed area offers more possible investigations than a six-month internship can cover, while the assignment is to one area. Prepare: Outline a bounded first study with a clear question, baseline and decision point for expanding or changing direction.
Distinguish promising results from reliable findings
Likely challenge (inference): Interpretability, uncertainty and robustness methods can look persuasive on selected examples yet perform differently across models or modalities. Prepare: Explain how you would test against baselines, examine failure cases and state the limits of any claim; the exact evaluation protocol is not specified.
Balance research iteration with maintainable outputs
Likely challenge (inference): Exploring new methods quickly can compete with the posting’s expectations for testable code, systematic studies and written research contributions. Prepare: Describe how you would keep experiments reproducible while reserving time to review, document and communicate results.

Practice your answers

Technical practice

Suggested practice: How do attention, positional encodings, tokenization and training objectives shape transformer behavior?

What to coverThe posting explicitly requires these fundamentals across architectures including BERT, GPT and T5. Compare how an architectural or training-objective choice affects what a model can learn and how it is evaluated.

Suggested practice: How would you diagnose a PyTorch model that improves on training data but not validation data?

What to coverThe posting requires ML and deep-learning foundations plus hands-on PyTorch experience. Discuss data leakage, overfitting and optimization as possible causes, and describe controlled checks rather than assuming one explanation.

Suggested practice: How would you structure an open-source Python library for a research method?

What to coverLibrary development and clean, modular, testable Python are stated parts of the role. Explain API boundaries, tests, documentation, packaging and a Git or CI workflow that helps research and engineering collaborators review changes.

Suggested practice: How would you evaluate an explanation method across text, images or tabular data?

What to coverThe posting names explainability work and methods such as SHAP, LRP and Grad-CAM. Choose a method appropriate to a modality, then discuss what its output does and does not establish and how you would test whether the explanation is useful.

Suggested practice: How could activation patching help investigate a model failure?

What to coverThe mechanistic-interpretability track explicitly mentions activation patching, circuits and feature visualization. Describe a comparison between successful and failing inputs, the intervention you would make and why the result would need careful interpretation.

Suggested practice: How would you benchmark uncertainty estimates alongside robustness?

What to coverThe posting names Bayesian approaches, ensembles, test-time augmentation and robustness metrics for foundation models. Compare the methods' computational costs and explain how you would check whether uncertainty remains informative when inputs change.

Suggested practice: How would you evaluate a proposed improvement to a tabular foundation model?

What to coverOne stated track covers tabular architectures and the TabTune library. Suggest appropriate baselines, dataset splits and repeatable comparisons, while distinguishing an architectural gain from a change in tuning or evaluation.

Suggested practice: What would you test when exploring an RL algorithm for a fine-tuning library?

What to coverThe posting identifies RL algorithms and an RL fine-tuning library as a possible track, without prescribing an algorithm. State the objective and baseline, then explain how you would compare learning behavior, stability and computational cost.

Suggested practice: How would you compare full-model fine-tuning with LoRA or adapters?

What to coverFull-model and parameter-efficient fine-tuning are listed as preferred domain experience, not universal requirements. Discuss trainable parameters, memory, adaptation quality and evaluation against the intended task.

Suggested practice: How would you design a systematic study whose results could support a research paper?

What to coverThe role includes maintaining experiment code, running systematic studies and contributing to whitepapers or submissions. Explain baselines, controlled comparisons, reproducibility and how you would report limitations or negative results.

Behavioural practice

Suggested practice: Tell me about a time you collaborated across research and engineering work.

What to coverThe posting calls for close collaboration between research and engineering teams. Use situation, action and result to show how you aligned goals, resolved a practical handoff and communicated the outcome.

Suggested practice: Describe a time code-review feedback changed your implementation.

What to coverPeer code reviews and collaborative development are explicitly mentioned. Give the situation, the feedback and your action, then explain the result for code quality or maintainability.

Suggested practice: Tell me about an experiment that did not support your initial hypothesis.

What to coverSystematic studies and research contributions make this a relevant preparation suggestion. Set out the hypothesis and situation, what you checked or changed, and how the result informed the next step.

Suggested practice: Describe how you communicated a technical finding to people with different backgrounds.

What to coverThe role calls for communication and contributions to research writing. Use situation, action and result to show how you explained the evidence, its limitations and its practical significance.

Suggested practice: Tell me about balancing speed with reliability in a technical project.

What to coverThe posting emphasizes both speed and reliability in hands-on research work. Describe the deadline or constraint, the checks you prioritized and the resulting trade-off without presenting a hypothetical achievement as your own.

Sharpen your knowledge

Topics to prepare

Machine learning and deep learning foundations
Required: the posting asks for a deep understanding of ML and deep learning principles. Revise by explaining how a model’s objective, optimization, and validation choices affect its results, then diagnose a hypothetical case of overfitting.
Model architectures and training objectives
Required: revise attention, positional encodings, tokenization, and training objectives. Sketch how BERT, GPT, LLaMA, T5, MoE, and Mamba differ, including where Mamba differs from attention-based transformers; these are named in the posting.
Interpretability methods
Track-specific preparation suggestion, not a requirement across all tracks: compare what SHAP, Grad-CAM, and activation patching can reveal, using one text or image prediction as a thought exercise. The posting lists explainability and mechanistic interpretability as possible research areas.
Uncertainty and robustness
Track-specific preparation suggestion: explain how ensembles and test-time augmentation estimate uncertainty, then propose a metric for comparing model behavior under a defined perturbation. Both methods and robustness benchmarking appear in one possible research area.
Systematic research studies
The role calls for systematic studies and research writing. Revise experimental design by outlining a baseline, one controlled change, evaluation metrics, and limitations for a chosen research question; the outline is a preparation exercise, not a stated employer protocol.

Skills to revise

Write testable Python research code
Required: practice splitting a small model-evaluation script into reusable functions and adding tests for input shapes and edge cases. This reflects the posting’s emphasis on clean, modular, testable Python and maintained experiment code.
Run and inspect a PyTorch experiment
Required: implement a small training loop, record training and validation results, and check gradients and tensor shapes. The posting explicitly requires hands-on PyTorch experience.
Prepare a collaborative code change
Required: make a focused Git change to the exercise code, document how to run it, and rehearse a peer review of the change. Git workflows, CI/CD, packaging, documentation, and code reviews are listed qualifications; adding a simple automated test run is a preparation suggestion.
Benchmark one chosen research method
Track-specific preparation suggestion: for an area of interest named in the posting, compare a method with a baseline using the same data split and metric, then summarize the result and a failure case. The intern would work on one research area, not all of them.
Communicate experimental findings
Draft a short methods-and-results summary that states the question, comparison, evidence, and limitations. This practices the communication skills and potential whitepaper or conference-submission contributions described in the posting.

Your preparation checklist

Choose a research focus
The posting says interns work on one research area, and the application asks for a preferred topic and a 1–2-year research direction. Prepare a brief explanation of your choice and one question you would investigate.
Prepare a project walkthrough
The application asks for a problem you are proud of solving. If you have relevant work, prepare a concise account of the problem, your contribution, PyTorch methods, results and limitations; otherwise, label a proposed example as hypothetical.
Review the required technical foundations
Prepare short explanations of attention, positional encodings, tokenization and training objectives, with comparisons among relevant model families such as BERT, GPT and T5. Be ready to explain a PyTorch implementation decision and how you would test it.
Draft an experiment for your chosen area
As interview preparation, outline a small study tied to one advertised area—for example, comparing uncertainty methods or evaluating an explainability technique. Specify a baseline, data, metrics, controls, failure cases and how results would be reproduced.
Gather evidence of research-ready code
The role calls for modular Python, maintained experiments, Git workflows, packaging and documentation. Identify an existing code sample or prepare a small demonstrator with tests, reproducible run instructions and a clear account of trade-offs.
Outline how you would communicate findings
The posting mentions systematic studies, peer review and possible whitepapers or submissions. Prepare a one-page results outline that separates observations from interpretations, notes limitations and identifies feedback you would seek from researchers and engineers.
Write questions for the team
Ask how the research area is selected, what an intern would own during the six-month term and how experiments are reviewed. Clarify the remote and Bangalore arrangements, since the posting says remote while the application asks for a preferred work location.

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…