New opportunity

Applied Scientist – Machine Learning

Recrew · Gurugram, India

About this role

This full-time, on-site Applied Scientist role in Gurugram, India, combines applied ML research and production engineering for an industrial AI startup’s Causal Intelligence Platform. The platform uses multimodal factory data and explainable causal reasoning to investigate equipment and process failures. The scientist will develop root-cause insights for engineers in semiconductor fabs, utilities and advanced manufacturing, with influence over scientific direction, product architecture and customer outcomes.

Responsibilities

Design ML algorithms for diagnostics using industrial time-series data, logs and waveforms.
Build and optimize LLM and RAG pipelines for structured extraction, hybrid search and domain-specific reasoning.
Architect scalable ingestion of documents, logs, sensor streams, alarms and images; work with engineers to turn ML models into reliable customer-facing systems.
Define use cases, benchmarks and evaluation frameworks with product teams. Lead projects from research and prototyping through deployment, iterating on customer feedback, and contribute to patents, technical architecture documents and research publications.

Required qualifications

BS, MS or PhD in AI/ML, Computer Science, Electrical Engineering, Applied Mathematics or a related field, plus 2–5 years of hands-on ML research or applied ML experience.
Strong Python skills and a record of building and scaling production ML systems. Hands-on experience with LLMs, RAG architecture, prompt engineering or fine-tuning; experience with time-series modeling, anomaly detection or statistical signal processing; and demonstrated ability to deploy ML projects as customer-facing systems.

Preferred qualifications include startup or early-stage product experience; knowledge of semiconductor, manufacturing, battery or industrial data; experience with agents or autonomous reasoning; open-source ML or LLM tooling contributions; and work with customers or field engineering teams. The posting offers ESOPs, growth opportunities and direct influence on product and scientific direction.

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

Skills for this role

PythonMachine LearningCausal AIMultimodal AITime-Series ModelingAnomaly DetectionStatistical Signal ProcessingLarge Language ModelsRAGPrompt EngineeringFine-TuningStructured Data ExtractionHybrid SearchInformation RetrievalData IngestionModel EvaluationProduction ML SystemsAI Agents

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

Industrial diagnostics and root-cause insights
The posting calls for ML algorithms using time-series data, logs and waveforms to produce explainable root-cause insights for engineers.
Domain-specific LLM and RAG pipelines
Build and optimize structured extraction, hybrid search and domain-specific reasoning over industrial information.
Multimodal ingestion and production reliability
Architect ingestion for documents, logs, sensor streams, alarms and images, then work with engineers to ship reliable customer-facing systems.
Evaluation tied to customer use cases
Define use cases, benchmarks and evaluation frameworks with product teams, and iterate on real customer feedback.
End-to-end scientific and technical ownership
Lead work from research and rapid prototypes through deployment, influence product architecture, and contribute to technical documents, patents or publications.

Likely challenges to think through

Distinguishing causes from correlations
Likely challenge, not a confirmed problem: industrial signals may support a plausible diagnostic without establishing a cause. Prepare to explain how you would test competing explanations, represent uncertainty and make conclusions useful to engineers.
Joining heterogeneous data without losing reliability
Likely challenge: documents, sensor streams, alarms and images differ in format and timing, while the posting calls for reliable systems in secure, private environments. Prepare to discuss ingestion design, data-quality checks and deployment trade-offs.
Balancing fast prototypes with trustworthy evaluation
Likely challenge: rapid iteration on customer feedback can pull against stable benchmarks and production readiness. Prepare to describe how you would choose measurable use cases, compare prototypes fairly and decide what is ready to ship.

Practice your answers

Technical practice

How would you detect equipment anomalies in industrial time-series data?

What to coverPreparation suggestion: Compare statistical or signal-processing baselines with learned models, accounting for changing operating conditions and limited failure labels. Explain how you would assess false alarms and whether a detected anomaly helps engineers investigate a failure.

How would you combine logs and waveforms to diagnose a process failure?

What to coverPreparation suggestion: Discuss aligning events and signals in time, handling missing or noisy observations, and identifying evidence that supports a diagnosis. Distinguish an observed association from a plausible cause.

How would you make a root-cause insight explainable rather than merely correlated?

What to coverPreparation suggestion: Describe what evidence from multimodal factory data would support or challenge a proposed causal explanation. Address uncertainty and how an engineer could verify the insight.

How would you extract structured facts from industrial documents and logs using an LLM?

What to coverPreparation suggestion: Explain your approach to defining fields, handling inconsistent source formats, and validating outputs against the original text. Discuss when prompt changes might suffice versus when further model adaptation could be useful.

How would you design hybrid search for a diagnostic RAG pipeline?

What to coverPreparation suggestion: Compare lexical and semantic retrieval for technical terminology, identifiers, and varied document language. Explain how you would evaluate whether retrieved evidence is relevant to an engineer’s question.

How would you reduce unsupported conclusions in domain-specific RAG answers?

What to coverPreparation suggestion: Discuss grounding answers in retrieved evidence, indicating uncertainty, and checking whether the evidence supports the stated reasoning. Explain how you would evaluate answer quality separately from retrieval quality.

How would you architect ingestion across sensor streams, alarms, images, logs, and documents?

What to coverPreparation suggestion: Address source-specific parsing, timestamps, provenance, data quality, and a representation that supports later diagnostics. Describe trade-offs between ingestion speed, reliability, and the ability to trace an insight back to its inputs.

How would you define benchmarks for a new industrial diagnostic use case?

What to coverPreparation suggestion: Start with the decision engineers need to make, then define representative cases, baselines, and measures for both diagnostic quality and practical usefulness. Discuss how you would avoid misleading results when failures are rare or labels are uncertain.

How would you turn a Python ML prototype into a reliable customer-facing system?

What to coverPreparation suggestion: Cover testing, monitoring, failure handling, and scaling in collaboration with engineering teams. Explain how deployment in a secure, private environment could affect design choices without assuming a particular infrastructure.

Behavioural practice

Tell me about an ML project you took from prototype to deployment.

What to coverPreparation suggestion: Use a real situation, your decisions and actions, and the observed result. Focus on what changed between research and a reliable user-facing system.

Describe a time customer or field feedback changed your technical approach.

What to coverPreparation suggestion: Set out the original assumption, the feedback you received, your response, and the result. Explain how you balanced fast iteration with evidence that the change improved the outcome.

Give an example of working with engineers or product partners to define an ML use case.

What to coverPreparation suggestion: Describe the situation, how you agreed on scope and evaluation criteria, what you did, and the result. Make your own contribution and the cross-functional trade-offs clear.

Tell me about a time you made progress on an ambiguous technical problem.

What to coverPreparation suggestion: Use a real example to show how you framed the uncertainty, prioritized experiments, and decided what to do next. State the result, including what you learned if the initial approach did not work.

Describe how you communicated a complex ML finding to people who needed to act on it.

What to coverPreparation suggestion: Explain the situation, how you presented evidence and limitations, and what action or decision followed. Choose an example that shows how you made a technical result understandable without overstating certainty.

Sharpen your knowledge

Topics to prepare

Industrial time-series diagnostics
The posting requires experience with time-series modeling, anomaly detection, or statistical signal processing and calls for diagnostics across sensor data, logs, and waveforms. Suggested revision: explain how you would distinguish a process shift from a sensor fault, including what evidence you would check before attributing a cause.
LLM and RAG design
Hands-on experience with LLMs, RAG architecture, prompt engineering, or fine-tuning is required; the work includes structured extraction, hybrid search, and domain-specific reasoning. Suggested revision: diagram a document-to-answer pipeline and identify where extraction, retrieval, and reasoning errors can occur.
Causal reasoning and explainability
The platform aims to provide explainable root-cause insights rather than surface-level correlations. Suggested revision: take a hypothetical equipment alarm and explain why an event preceding a failure is not necessarily its cause, then identify evidence that would strengthen the diagnosis. This is preparation for the product context, not a separately stated qualification.
Benchmarks for root-cause insights
The role defines use cases, benchmarks, and evaluation frameworks with product teams. Suggested revision: specify what a benchmark would label, how it would separate historical cases for evaluation, and how it would assess both diagnostic accuracy and the usefulness of supporting evidence.
Research-to-deployment trade-offs
The posting calls for leading ML projects from experimentation through reliable customer-facing deployment, with solutions used in secure, private environments. Suggested revision: compare a promising prototype with a deployable version, covering failure handling, reproducibility, and how customer feedback would change the next iteration.

Skills to revise

Implement a Python diagnostic baseline
Strong Python and applied time-series experience are required. As a suggested exercise, write a small Python program that aligns timestamped sensor readings with alarms, flags unusual intervals, and outputs the readings supporting each flag; test missing and out-of-order data.
Prototype extraction and retrieval
The role builds LLM/RAG pipelines for structured extraction and hybrid search. As a suggested exercise, extract fields from a few sample incident documents, retrieve relevant passages for a diagnostic question, and inspect where the answer lacks support in the retrieved material.
Design multimodal ingestion
The role architects ingestion for documents, logs, sensor streams, alarms, and images. As a suggested exercise, sketch a common event record with source, timestamp, equipment identifier, and provenance, then trace how an image and its related alarm would remain linked. The record design is a preparation suggestion, not a prescribed company schema.
Run a focused model evaluation
The role defines evaluation frameworks and optimizes ML pipelines. As a suggested exercise, assemble a small set of hypothetical fault cases, record expected evidence and root cause, compare diagnostic outputs against them, and categorize misses such as unsupported explanations or overlooked alarms.
Plan a production handoff
The posting requires experience scaling production ML systems and deploying customer-facing projects with engineering teams. As a suggested exercise, write a one-page handoff for a diagnostic prototype covering inputs, outputs, test cases, failure behavior, and what feedback engineers should capture after deployment.

Your preparation checklist

Map your evidence to the stated requirements
Prepare a one-page list of examples you can substantiate for Python, production ML, LLM or RAG work, and time-series modeling, anomaly detection or signal processing. Note relevant education and applied-ML experience without filling gaps with unsupported claims.
Rehearse an end-to-end ML project walkthrough
If you have a relevant project, outline the problem, your decisions, experiments, deployment, scaling, reliability and measured results. The role calls for taking ML work from research to a customer-facing system.
Work through an industrial diagnostics example
As a practice exercise, sketch how you would align sensor time-series, logs and waveforms, detect an abnormal event, test possible causes and explain uncertainty to an engineer. Distinguish a causal hypothesis from a correlation.
Draft a grounded RAG pipeline design
Prepare a diagram showing document and log extraction, indexing, hybrid search, domain-specific reasoning and answer evaluation. Be ready to explain how you would check retrieved evidence and handle unsupported conclusions.
Sketch a secure multimodal ingestion architecture
As an interview exercise, show how documents, sensor streams, alarms and images could enter a scalable ML system. Identify data-quality, monitoring and deployment trade-offs, given the posting’s emphasis on reliable systems in secure, private environments.
Prepare an evaluation plan
Write down proposed baselines and metrics for diagnostics and retrieval, plus how product and engineering teams could review results and customer feedback. The posting explicitly assigns the scientist a role in defining use cases, benchmarks and evaluation frameworks.
Bring questions about scientific and product priorities
Ask which failure modes and data types matter most, what evidence makes a root-cause insight useful to engineers, and how the team evaluates prototypes before customer deployment. These are questions to explore, not assumptions about current 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…