New opportunity

LLM Research Engineer

Predii · Pune, India

About this role

Predii seeks an LLM Research Engineer in Pune with 1–4 years of experience to research and develop models for its platform and apply that work to the automotive ecosystem. The role focuses on question answering, fine-tuning with synthetic and proprietary datasets, and turning model capabilities into back-end services.

Responsibilities

Own the design and development of algorithms, and maintain and improve Predii's LLM models and supporting software.
Build services using microservices, REST APIs, ML and NLP libraries, customized algorithms, and NoSQL datastores.
Develop libraries and frameworks for large, complex web applications, contributing from planning through execution to efficient, scalable, distributed solutions.
Work as an individual contributor with cross-functional teams, customers, and open-source communities in Predii Labs; participate in LLM research and development of new services.

The posting lists Computer Vision, Generative AI, Machine Learning, NLP, Python, PyTorch, and Scikit-learn among relevant skills. Predii's platform works with unstructured textual, sensor, and procedural data in the automotive aftermarket and service business.

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

Skills for this role

Large Language ModelsQuestion AnsweringFine-TuningSynthetic DataGenerative AIMachine LearningNatural Language ProcessingComputer VisionPythonPyTorchScikit-learnMicroservicesREST APIsNoSQLAlgorithm DesignDistributed 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

Improve LLM question answering
The role builds and enhances Predii’s LLM models for question answering and owns algorithm design and model maintenance. Prepare to explain how you would evaluate an answer-quality improvement and decide whether to adopt it.
Fine-tune with synthetic and proprietary data
Fine-tuning on both dataset types is a stated daily duty. Prepare to discuss how you would check data quality, compare model results, and investigate errors without assuming the two sources behave alike.
Turn models into back-end services
The posting calls for productizing model work with microservices, REST APIs, ML and NLP libraries, customized algorithms, and NoSQL datastores. Prepare an example of moving a model capability into a maintainable service.
Build for the automotive domain at scale
Research must be applied to the automotive ecosystem, while engineering work spans planning through execution of scalable, distributed solutions. Predii describes working with unstructured service data; prepare to explain how domain-specific data could affect model and system design.

Likely challenges to think through

Balancing research gains against service constraints
A likely challenge, not a confirmed Predii problem, is that a better question-answering model may cost more to run or be harder to serve reliably. Prepare to describe how you would compare answer quality with latency, resource use, and implementation complexity before productizing a change.
Learning from uneven domain data
Synthetic and proprietary datasets, alongside the unstructured automotive data Predii describes, may differ in coverage and terminology. Prepare to explain how you would test whether fine-tuning improvements hold on representative domain examples rather than only on training-like data.
Owning work across team boundaries
The role combines individual ownership with work alongside cross-functional teams, customers, and open-source communities. A likely difficulty is reconciling research ideas with service needs and external feedback. Prepare an example of communicating a technical trade-off and carrying a decision through implementation.

Practice your answers

Technical practice

How would you evaluate an LLM used for question answering?

What to coverThe role explicitly involves building and enhancing question-answering models. Suggested preparation: explain how you would test answer correctness and relevance on representative questions, inspect failure cases, and compare improvements against a baseline.

How would you approach fine-tuning with synthetic and proprietary datasets?

What to coverBoth dataset types are named in the posting. Suggested preparation: discuss data selection, train/evaluation separation, and how you would check whether fine-tuning improves the intended task rather than merely fitting the training examples.

How would you assess the quality of synthetic training data?

What to coverSynthetic data is explicitly part of the fine-tuning work. Suggested preparation: describe checks for inaccurate or repetitive examples and how you would decide whether the added data improves performance on independently evaluated questions.

How would you adapt question answering to automotive service data?

What to coverPredii describes working with unstructured textual, sensor, and procedural data in automotive servicing. Suggested preparation: explain how you would identify the information needed to answer a service question and handle ambiguous or incomplete records without overstating certainty.

How would you design and improve an algorithm used by an LLM-backed service?

What to coverThe role assigns ownership of algorithm design and ongoing model and software improvements. Suggested preparation: start with a measurable problem, compare a simple baseline with an alternative, and explain how evaluation and maintenance needs would affect your choice.

How would you expose a model capability through a REST microservice?

What to coverThe posting calls for productizing back-end services with microservices and REST APIs. Suggested preparation: cover request and response design, input validation, error handling, and the trade-off between model response time and service reliability.

What would you consider when using a NoSQL datastore for this work?

What to coverNoSQL datastores are named among the back-end components. Suggested preparation: discuss access patterns, data representation, and how retrieval needs and changing records would influence a datastore design; avoid assuming a particular database.

How would you make a model-backed service scalable and maintainable?

What to coverThe responsibilities mention libraries and frameworks for complex web applications and efficient, scalable, distributed solutions. Suggested preparation: explain how you would separate reusable model code from service logic, locate performance bottlenecks, and test changes before scaling them.

How would you test whether an LLM research idea is ready for a service?

What to coverThe role combines LLM research with development of new services; Python, PyTorch, and Scikit-learn appear in the skills list. Suggested preparation: describe a reproducible experiment, a comparison with an existing approach, and the practical trade-offs you would review before implementation.

Behavioural practice

Tell me about a time you took ownership of a technical improvement.

What to coverThe role expects an individual contributor to own algorithm and model improvements. Suggested preparation: use a real situation, specify what you decided and implemented, and describe the observed result without claiming work you did not do.

Tell me about a time you worked across disciplines to solve a problem.

What to coverThe posting names cross-functional collaboration. Suggested preparation: describe the situation, how you aligned differing technical or product needs, your own actions, and the result.

Tell me about a time customer input changed your technical approach.

What to coverWorking with customers is explicitly part of the role. Suggested preparation: choose a real example showing what you learned, what you changed, and what happened; distinguish direct customer contact from feedback relayed by someone else.

Tell me about a time you contributed to or learned from an open-source community.

What to coverThe role includes work with open-source communities. Suggested preparation: describe a real interaction or contribution, the actions you took to evaluate or communicate it, and its outcome; do not imply a contribution if your experience was limited to learning.

Tell me about a time you moved an experiment toward a usable solution.

What to coverThe posting connects research with new services and engineering from planning through execution. Suggested preparation: explain the initial uncertainty, the decisions and collaboration involved, and what the resulting implementation or experiment established.

Sharpen your knowledge

Topics to prepare

LLM question answering
Posting: The role builds and improves LLM models for question answering. Revision exercise: explain how you would judge whether an answer is supported by the input and compare two possible answers to a sample automotive service question. Using that comparison as an evaluation exercise is a preparation suggestion, not a stated employer method.
Fine-tuning with synthetic and proprietary data
Posting: Fine-tuning uses synthetic and proprietary datasets. Revision exercise: outline how you would prepare examples from each type of data and check whether a model improved on question answering. The checks are suggested preparation, not a specified Predii process.
Automotive service-data context
Posting: Predii works with unstructured textual, sensor, and procedural data; its platform includes algorithms tuned to complaint, cause, and correction. Revision exercise: take a fictional repair record and identify those three elements, then explain what an answering model could and could not conclude from them.
Research-to-service architecture
Posting: The role connects LLM research to back-end services using microservices, REST APIs, and NoSQL datastores, while contributing to scalable, distributed solutions. Revision exercise: sketch the components and data flow for a question-answering service and explain where model updates and datastore access would fit. The design is a practice scenario, not Predii's stated architecture.

Skills to revise

Python and PyTorch model work
Posting: Python and PyTorch are listed skills, and the role maintains and improves LLM models. Revision exercise: write a small PyTorch fine-tuning loop using mock question-answer examples and inspect training outputs; mock data avoids assuming access to Predii datasets.
ML and NLP comparison with Scikit-learn
Posting: ML, NLP, and Scikit-learn are listed skills. Revision exercise: use a small sample of service-text questions to build a simple Scikit-learn comparison and record where its answers differ from an LLM's. This comparison is suggested practice, not a required Predii approach.
REST service implementation
Posting: Productizing back-end services with microservices and REST APIs is part of the role. Revision exercise: implement a small Python endpoint that accepts a question, calls a stubbed answer function, and returns a structured response; test malformed inputs as well as valid ones.
Algorithm and library maintenance
Posting: The individual contributor owns algorithm design and helps build libraries and frameworks for complex web applications. Revision exercise: package a small text-processing algorithm as a reusable Python function, add tests for ambiguous inputs, and explain one improvement you would make after reviewing failures.

Your preparation checklist

Prepare an LLM project walkthrough
The role involves owning algorithms and improving LLM models. Prepare a concise walkthrough of a project you can discuss, covering the problem, your individual contribution, experiments, evaluation, limitations, and what you changed.
Practice a question-answering design exercise
Predii builds LLMs for question answering and applies research to automotive service data. As interview practice, sketch a system that answers questions from unstructured repair text; show how you would test answer quality and handle uncertain answers.
Outline a fine-tuning experiment
The posting names synthetic and proprietary datasets. Prepare an experiment plan using data you are permitted to discuss: dataset preparation, a baseline, a fine-tuning approach, evaluation, and checks for misleading synthetic examples.
Map the model to a back-end service
The role productizes models with microservices, REST APIs, and NoSQL datastores. Sketch an API and data flow for serving answers, then explain likely scaling, latency, failure-handling, and maintenance trade-offs as design considerations.
Select evidence for the listed tools and collaboration
Identify concrete examples, where available, of your Python, PyTorch, Scikit-learn, or NLP work. Prepare to explain how you worked with others or incorporated feedback; Predii describes an individual-contributor role involving cross-functional teams, customers, and open-source communities.
Draft focused questions for Predii
Ask how Predii evaluates question-answering improvements, how research moves into back-end services, and which automotive data problems the team is prioritizing. Treat these as questions, not assumptions about its processes.

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…