New opportunity

NLP Engineer

Zimmer Biomet · Bangalore, India

About this role

Zimmer Biomet is hiring an NLP Engineer for its Enterprise AI team in Bangalore, India, operating on a hybrid schedule of 2 days work from home and 3 days in the office. Reporting to the Director of AI, this delivery-focused role designs, builds, evaluates, and deploys production natural language processing and large language model systems for regulatory, clinical, quality, commercial, and support text across global business functions.

Responsibilities

  • Design and build retrieval-augmented generation (RAG) pipelines, covering chunking strategy, embedding selection, vector storage, retrieval tuning, and reranking.
  • Develop, version, test, and maintain prompt and context engineering artifacts, as well as multi-step agent and tool-calling workflows with failure handling.
  • Build evaluation harnesses for generative systems measuring groundedness, factual accuracy, hallucination rate, refusal behavior, latency, and cost per request.
  • Work with Azure OpenAI and other approved model providers to balance task quality, cost, and latency.
  • Build document extraction pipelines converting long, unstructured, and inconsistently formatted files (including scanned PDFs, OCR output, tables, and multi-column layouts) into structured data using named entity recognition, classification, summarization, and relation extraction.
  • Define extraction metrics (precision, recall, error analysis on held-out sets) and implement human-in-the-loop review steps where confidence thresholds or regulatory rules require.
  • Select, fine-tune, and evaluate language and embedding models, manage semantic search indexes, curate labeled datasets with annotation guidelines and inter-annotator agreement, establish baselines, and conduct error analysis.
  • Write tested, reviewed, documented production Python code, deploy services on Microsoft Azure with CI/CD and observability, monitor quality drift, cost, and latency, and enforce data privacy, information security, and responsible AI policies.

Required qualifications

  • Bachelor's or Master's degree in computer science, computational linguistics, data science, engineering, mathematics, or a related field, or equivalent practical experience.
  • 3 to 5 years of experience building NLP or machine learning systems, including at least one deployed to production for real users.
  • Demonstrable experience building LLM applications in production, including RAG and prompt engineering.
  • Experience extracting structured data from unstructured documents at scale and defining quantitative model or pipeline evaluation criteria.
  • Experience partnering with non-technical stakeholders across time zones to scope ambiguous business problems and communicate model behavior and limitations in English.
  • Strong proficiency in Python (PyTorch or TensorFlow, Hugging Face Transformers, spaCy, scikit-learn, pandas), LLM frameworks (LangChain, LlamaIndex, or equivalent), vector databases (Azure AI Search, FAISS, pgvector, or Pinecone), transformer architectures, tokenization, embeddings, fine-tuning, SQL, Git, automated testing, CI/CD, and cloud deployment (ideally Microsoft Azure).

Preferred qualifications

  • Experience in a regulated industry (medical devices, pharmaceuticals, or healthcare) and familiarity with regulatory documentation expectations.
  • Experience with clinical, regulatory, or quality text, or terminologies such as SNOMED, MedDRA, or ICD.
  • Experience with document AI and OCR tools such as Azure Document Intelligence.
  • Experience with MLOps practices (model registries, experiment tracking, automated retraining or re-evaluation) and responsible AI governance (bias evaluation, model cards).
  • Contributions to open-source NLP tooling, publications, or comparable public work, and experience in a Global Capability Center supporting global business functions.

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

Skills for this role

PythonPyTorchTensorFlowHugging Face TransformersspaCyscikit-learnpandasLangChainLlamaIndexAzure AI SearchFAISSpgvectorPineconeSQLGitCI/CDMicrosoft AzureAzure OpenAIAzure Document IntelligenceRetrieval-Augmented GenerationPrompt EngineeringNamed Entity RecognitionOCRMLOpsFine-tuningEmbeddingsSemantic Search

Your skill match

Checking your profile…

YOUR NEXT STEP

Get interview-ready for this role

A focused preparation guide, built around this job’s responsibilities and requirements.

✦ AI-generated guide
Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.
01

Understand the role

What matters most

  • End-to-End Production RAG and Agentic Pipelines

    Source-backed priority: Design, build, and deploy retrieval-augmented generation systems (covering chunking, embedding selection, vector storage, retrieval tuning, and reranking), versioned prompt engineering artifacts, and multi-step agentic or tool-calling workflows using Azure OpenAI and approved providers.

  • Structured Extraction from Complex Unstructured Documents

    Source-backed priority: Build pipelines converting scanned PDFs, OCR output, multi-column layouts, tables, and inconsistent clinical, regulatory, quality, and commercial documents into structured data via named entity recognition, classification, summarization, and relation extraction, supported by human-in-the-loop review where required.

  • Rigorous Quantitative Evaluation and Error Analysis

    Source-backed priority: Because the company operates in a regulated industry where correctness outweighs novelty, the role prioritizes building evaluation harnesses for groundedness, factual accuracy, hallucination rate, refusal behavior, precision, recall, latency, and cost per request, alongside establishing simple baselines and conducting systematic error analysis.

  • Production Engineering, Observability, and Compliance on Azure

    Source-backed priority: Ship maintainable, tested Python services to production on Microsoft Azure with CI/CD, monitor deployed systems for quality drift, cost, and latency, and enforce data privacy, information security, responsible AI, and auditability documentation.

  • Cross-Regional Stakeholder Scoping and Generalizable Delivery

    Source-backed priority: Partner directly with non-technical business stakeholders across time zones (such as quality engineers, regulatory writers, and commercial teams) to translate ambiguous problems into scoped technical solutions that can be generalized across multiple business units.

Likely challenges to think through

  • Balancing Generative Flexibility Against Regulated Medical Device Accuracy

    Preparation hypothesis: Operating in a regulated medical technology environment means hallucinations or silent extraction errors in regulatory submissions or complaint records carry high compliance risk. Prepare concrete examples of how you enforce strict groundedness, design refusal behaviors, set confidence thresholds for human-in-the-loop escalation, and document failure modes transparently for audits.

  • Extracting Reliable Data from Noisy, Multi-Layout Enterprise Documents

    Preparation hypothesis: Clinical evaluation reports, surgical technique guides, and scanned PDFs often break naive chunking and parsing strategies due to complex tables and multi-column layouts. Be ready to discuss layout-aware OCR and document parsing trade-offs, held-out error analysis workflows, and how you combine classical NLP (NER, relation extraction) with LLMs rather than relying solely on generative prompts.

  • Justifying Model Complexity Against Latency, Cost, and Simple Baselines

    Preparation hypothesis: The posting explicitly emphasizes establishing baselines first and balancing quality against cost and latency across approved providers. Prepare to explain how you decide between lightweight classical models, fine-tuned embeddings, single-call RAG, and multi-step agent workflows, backing architectural choices with quantitative evaluation metrics rather than qualitative demos.

  • Managing Parallel Use Cases Across Distributed Global Stakeholders

    Preparation hypothesis: Supporting multiple global business functions from a Bangalore hub (3 days in office, 2 days WFH) requires aligning expectations across time zones and generalizing pipelines across units. Prepare stories demonstrating how you translate vague business requests into measurable success criteria, push back when AI is not the right fit, and modularize pipelines for reuse.

03

Practice your answers

Technical practice

How would you design a chunking, retrieval, and reranking strategy for long, multi-column surgical technique guides and regulatory submissions?

What to coverSuggested practice question: Address layout-aware parsing (preserving section headers, tables, and multi-column reading order) before chunking rather than naive fixed character splits. Compare hybrid retrieval (BM25/keyword plus dense embeddings in stores like Azure AI Search or pgvector) and explain why a cross-encoder reranker improves precision at the cost of added latency.

How do you build an automated evaluation harness to measure groundedness, hallucination rate, and refusal behavior in a production RAG system?

What to coverSuggested practice question: Detail separating retrieval metrics (recall@k, MRR) from generation metrics (claim-level faithfulness/groundedness against retrieved context, factual accuracy, and appropriate abstention on unanswerable queries). Discuss using deterministic checks, held-out golden datasets, and calibrated LLM-as-a-judge evaluators while tracking latency and cost per request.

When would you choose classical NLP or fine-tuned encoder models over a generative LLM for structured document extraction?

What to coverSuggested practice question: Contrast deterministic/encoder approaches (spaCy, Hugging Face NER/classification) with LLM-based extraction across accuracy, schema reliability, latency, token cost, and auditability. Emphasize starting with a simple baseline on a held-out set and only introducing generative or fine-tuned models when error analysis proves the simpler baseline fails.

How would you handle noisy scanned PDFs, complex tables, and inconsistent templates in an extraction pipeline requiring human-in-the-loop review?

What to coverSuggested practice question: Explain combining OCR/layout extraction (such as Azure Document Intelligence) with schema validation, field-level confidence calibration, and rule-based sanity checks. Describe how low-confidence or regulatory-sensitive extractions route to human reviewers and how corrected outputs feed back into error analysis.

What trade-offs govern embedding model selection, tokenization limits, and fine-tuning versus prompt/retrieval optimization?

What to coverSuggested practice question: Cover context-window truncation risks, domain terminology handling in tokenizers, and vector dimensionality vs. storage/search speed. Explain why improving retrieval quality and prompt/context structure usually precedes embedding or LLM fine-tuning, and what labeled data volume and maintenance overhead fine-tuning introduces.

How do you design a multi-step agent or tool-calling workflow so that failures, loops, and bad tool outputs are safely contained?

What to coverSuggested practice question: Focus on clear task decomposition, strict input/output schemas for tools, bounded step/retry budgets, and graceful degradation or refusal when external calls fail. Reinforce the posting's pragmatism principle by explaining how you verify a single-call or deterministic pipeline is insufficient before adding agentic complexity.

How do you structure, version, test, and deploy prompt and context engineering artifacts within a production CI/CD pipeline on Azure?

What to coverSuggested practice question: Describe storing prompts as versioned code artifacts in Git rather than hardcoded strings, paired with automated regression tests against a benchmark dataset on pull requests. Discuss containerizing the Python service, deploying via CI/CD on Microsoft Azure, and isolating staging vs. production model endpoints.

How do you monitor a deployed NLP/LLM service for quality drift, latency spikes, and cost anomalies over time?

What to coverSuggested practice question: Outline logging structured telemetry (token counts, per-step latency, retrieval similarity distributions, refusal rates, and user/reviewer feedback) without leaking sensitive data. Explain setting alert thresholds and running periodic scheduled re-evaluations on curated slices to catch upstream model updates or document template drift.

How do you design dataset annotation guidelines and measure inter-annotator agreement when building a ground-truth evaluation set?

What to coverSuggested practice question: Explain writing unambiguous edge-case rules for domain annotators, running pilot calibration rounds, and quantifying agreement using metrics like Cohen's kappa or span-level F1 for NER. Show how resolving disagreements improves both label quality and downstream error analysis.

Behavioural practice

Tell me about a time you reported a negative evaluation result or model limitation instead of pushing forward a flattering demo.

What to coverSuggested practice question (Situation/Action/Result): Describe the context where initial metrics or stakeholder expectations looked promising, how your quantitative error analysis uncovered critical failure modes, and how you communicated those limits transparently to realign scope or fix the root cause.

Describe a situation where you had to push back on unnecessary technical complexity and advocate for a simpler baseline solution.

What to coverSuggested practice question (Situation/Action/Result): Walk through a problem where stakeholders or teammates wanted a complex LLM or agentic architecture, how you benchmarked a simpler baseline first, and what trade-offs in accuracy, latency, cost, or maintainability justified your final recommendation.

Give an example of partnering with non-technical stakeholders to turn an ambiguous request into a scoped, measurable NLP deliverable.

What to coverSuggested practice question (Situation/Action/Result): Explain how you elicited the underlying workflow bottleneck, translated vague goals into concrete precision/recall or latency acceptance criteria, managed expectations on feasibility, and delivered a production system.

Tell me about a time you diagnosed and resolved a subtle quality issue in a production NLP or machine learning system.

What to coverSuggested practice question (Situation/Action/Result): Detail how the issue surfaced (monitoring drift, user reports, or audit checks), how you isolated whether the root cause was data/OCR quality, retrieval, prompt logic, or model behavior, and the targeted iteration you shipped to resolve it.

Describe how you have managed multiple parallel use cases across distributed teams or time zones while maintaining code and documentation standards.

What to coverSuggested practice question (Situation/Action/Result): Share how you prioritized competing stakeholder requests, generalized reusable components across business units, and maintained clear design and evaluation documentation so others could review and build on your work.

02

Sharpen your knowledge

Topics to prepare

  • Transformer Architectures, Tokenization, Embeddings, and Fine-Tuning Trade-Offs

    Explicit requirement: Solid understanding of transformer internals, tokenization, embedding selection, and fine-tuning trade-offs. Revision exercise: Compare bi-encoder vs. cross-encoder reranking and full fine-tuning vs. parameter-efficient fine-tuning (such as LoRA) on domain-specific clinical or regulatory text, mapping out memory, latency, data volume, and maintenance trade-offs.

  • Retrieval-Augmented Generation (RAG) Architecture and Retrieval Tuning

    Explicit requirement: Design and build RAG pipelines covering chunking strategy, embedding selection, vector storage, retrieval tuning, and reranking. Revision exercise: Diagram an end-to-end RAG system for multi-column technical manuals or regulatory submissions, contrasting fixed-size, semantic, and layout-aware chunking strategies alongside hybrid (keyword plus dense vector) search and reranking metrics.

  • Quantitative Evaluation Harnesses for Generative and Extraction Systems

    Explicit requirement: Build evaluation harnesses measuring groundedness, factual accuracy, hallucination rate, refusal behavior, latency, and cost per request, plus precision, recall, and inter-annotator agreement on held-out sets. Revision exercise: Design an evaluation rubric and automated test suite for a document Q&A and entity extraction workflow, specifying how you compute Cohen's kappa for annotation guidelines and detect quality drift in production.

  • Multi-Step Agentic and Tool-Calling Workflows with Failure Handling

    Explicit requirement: Build agent and tool-calling workflows where a single model call is insufficient, including task decomposition, tool integration, and failure handling, while favoring the simplest approach that meets requirements. Revision exercise: Walk through a multi-step complaint-triage workflow that decomposes user queries, invokes structured search tools, validates outputs against schemas, and defines deterministic fallback or human-in-the-loop escalation rules.

  • Data Privacy, Information Security, and Responsible AI in Regulated Environments

    Explicit requirement (with preferred medical device/healthcare context): Handle sensitive data under privacy and security policies, flag validation obligations, and document design decisions and known limitations for audits; preferred familiarity with bias evaluation, model cards, and governance frameworks. Revision exercise: Draft a concise model card and audit checklist for a regulated clinical document extraction pipeline, covering PII redaction, refusal thresholds, and known failure modes.

Skills to revise

  • Production RAG and Vector Retrieval Implementation

    Explicit requirement: Practical command of Python, LangChain or LlamaIndex, and vector databases (Azure AI Search, FAISS, pgvector, or Pinecone). Revision exercise: Build a reproducible Python script using LangChain or LlamaIndex with FAISS or pgvector that ingests multi-section PDFs, runs hybrid retrieval with a reranker, and logs retrieval recall@k and per-query latency.

  • Structured Information Extraction and OCR Pipeline Engineering

    Explicit requirement (plus preferred Azure Document Intelligence): Extract structured data from scanned PDFs, OCR output, tables, and multi-column layouts using NER, classification, and relation extraction with spaCy, Hugging Face Transformers, or pandas. Revision exercise: Create a Python pipeline that parses a noisy multi-column PDF containing tables, runs schema-validated extraction or NER, and evaluates precision and recall against a small manually annotated JSON ground-truth set.

  • Versioned Prompt Engineering and Automated Evaluation Testing

    Explicit requirement: Treat prompts as versioned, tested artifacts with automated testing, Git, and CI/CD. Revision exercise: Set up a Git-tracked prompt template repository paired with a pytest evaluation suite that runs sample inputs against an LLM API (such as Azure OpenAI), checking JSON schema adherence, groundedness assertions, refusal behavior, and token cost per request.

  • Classical NLP and Fine-Tuning Baselines with PyTorch/Hugging Face

    Explicit requirement: Strong Python with PyTorch or TensorFlow, Hugging Face Transformers, spaCy, and scikit-learn, plus establishing baselines before adding complexity. Revision exercise: Train a fast TF-IDF + scikit-learn logistic regression baseline on a multi-label text classification dataset, then fine-tune a compact Hugging Face Transformer in PyTorch and compare F1 score, inference latency, and error categories.

  • Cloud Service Packaging, CI/CD, and Observability on Microsoft Azure

    Explicit requirement (plus preferred MLOps): Deploy and run NLP services on Microsoft Azure rather than notebooks, working with SQL, CI/CD, and monitoring for quality drift, cost, and latency. Revision exercise: Package a FastAPI NLP inference endpoint in Docker, write a CI workflow running unit and evaluation tests, and define structured logging queries to track token usage, p95 latency, and confidence-score distributions.

04

Your preparation checklist

  • Prepare a Production NLP/LLM System Walkthrough

    Document one end-to-end NLP or LLM system you shipped to real users (matching the 3-5 year production requirement), covering problem framing with non-technical stakeholders, baseline comparison, architecture, and post-launch monitoring.

  • Draft a Regulated Document Extraction and RAG Architecture Brief

    Create a one-page design exercise tailored to Zimmer Biomet's stated use cases (complaint records, regulatory submissions, or surgical technique guides), detailing OCR/layout parsing, hybrid vector retrieval, reranking, and human-in-the-loop confidence thresholds.

  • Assemble Quantitative Evaluation and Error Analysis Evidence

    Prepare concrete examples showing how you defined held-out metrics (precision, recall, groundedness, hallucination rate, refusal behavior, latency, and cost per request), conducted error analysis, and reported negative results transparently rather than relying on demos.

  • Review Azure Production Deployment and MLOps Artifacts

    Be ready to explain how you package Python services beyond notebooks using Git, automated tests, CI/CD, and cloud services (ideally Microsoft Azure, Azure OpenAI, Azure AI Search, or Azure Document Intelligence), plus how you track quality drift and token costs.

  • Outline Two Cross-Functional Stakeholder Stories

    Write structured Situation/Action/Result notes for two scenarios: translating an ambiguous business request into measurable technical scope, and communicating model limitations or regulatory/privacy constraints clearly to non-technical teams across time zones.

  • Prepare Targeted Questions for the Enterprise AI Team

    Draft 3-4 specific questions about how Enterprise AI prioritizes use cases across global business functions, how quality and regulatory teams validate NLP outputs, and how shared Azure infrastructure is managed with platform engineers.

AI-generated from the job information · Prepared Nov 15, 2026 (UTC)Useful preparation, not a guarantee of eligibility or an offer.

Availability has not yet been verified. 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…