NLP engineer

What does a NLP engineer do?

An NLP engineer builds software that reads, sorts, searches and writes human language.

The work covers chatbots, document search, text classification, summarisation and speech-to-text pipelines.

In Singapore, NLP engineers sit in AI or data teams in banks, insurers, government agencies, telcos, e-commerce firms and start-ups. The role fits the SSOC 2024 group for artificial intelligence and machine learning engineers.

You work with data scientists, data engineers, product managers and software developers. You also work with compliance, legal and business staff who know the documents and customer questions the system must handle.

What the work involves

• Clean and prepare text data, such as call transcripts, contracts, emails and chat logs.
• Train or fine-tune language models with PyTorch and Hugging Face, then compare results against a baseline.
• Build retrieval pipelines that connect an LLM API to a vector database so answers come from company documents.
• Test model output for errors, bias and unsafe replies, and fix the bugs you find.
• Deploy models as services, then monitor accuracy, speed and cost once real users arrive.

Pay in Singapore

Back to top

Skills a NLP engineer needs

Back to top

NLP engineer job description

We are hiring an NLP engineer to build and run language-based AI features for our Singapore teams.

You will turn text and speech data into working products, such as search, chat assistants and document tools. You will work with product, data and compliance colleagues, and you will own models from first experiment to live service.

Responsibilities

• Build NLP pipelines for tasks such as classification, entity extraction, summarisation and question answering.
• Design retrieval-augmented generation (RAG) systems using LLM APIs and vector databases such as Pinecone or pgvector.
• Fine-tune and evaluate open models from Hugging Face on company data.
• Prepare, label and version text datasets, and check them for quality and bias.
• Write evaluation sets and track measures such as accuracy, hallucination rate, latency and cost.
• Deploy models as APIs and work with MLOps or platform teams on monitoring and updates.
• Find and fix bugs, drift and poor answers in live models.
• Handle personal data in line with the Personal Data Protection Act (PDPA) and company policy.
• Explain model limits and results to non-technical colleagues in everyday language.

Requirements

• A degree in computer science, data science, linguistics, engineering or a related field. Strong self-taught or bootcamp candidates with a portfolio are also considered.
• Typically 1 to 4 years of experience in machine learning or NLP for mid-level roles. Junior roles may accept graduates with solid projects.
• Strong Python, plus experience with PyTorch and the Hugging Face libraries.
• Hands-on use of LLM APIs, prompt design, embeddings and vector databases.
• Working knowledge of Git, SQL, Docker and a cloud platform. Experience with English plus Mandarin, Malay or Tamil text is a plus.

Back to top

NLP engineer career path

NLP engineers usually start in junior machine learning or data roles and move towards technical leadership or AI management.

Titles vary by employer.

1. Junior NLP engineer

   Entry

   You clean text data, run experiments and fix bugs under guidance. You learn the team's tools and review practices.

2. NLP engineer

   Mid

   You own a model or feature from design to deployment. You choose methods, set up evaluation and work directly with product teams.

3. Senior NLP engineer

   Senior

   You lead the technical design of larger systems, such as enterprise RAG platforms. You review the work of others and mentor junior engineers.

4. Lead or principal NLP engineer

   Lead

   You set technical direction across several projects. You weigh cost, risk and quality, and you advise managers on what is feasible.

5. Head of AI or head of NLP

   Head

   You run the AI team and its budget. You answer to senior management for results, risk and governance.

The titles are ones commonly used in Singapore; employers may name the levels differently.

Related jobs and their pay

Back to top

NLP engineer interview questions

We wrote these questions from the skills this job needs. They are not an official question bank.

1. How does a transformer model handle context in a sentence, and why did it replace recurrent networks for most NLP tasks?

   What a good answer covers

   • Explains self-attention as weighing every token against the others
   • Notes that attention runs in parallel, while recurrent networks process tokens in sequence
   • Mentions better handling of long-range dependencies
   • Names a limit, such as cost growing with input length

2. You need to sort customer emails by topic, with 2,000 labelled examples across about 20 classes. What approaches would you compare?

   What a good answer covers

   • Starts with a simple baseline such as TF-IDF with logistic regression
   • Considers fine-tuning a small pretrained model from Hugging Face
   • Considers few-shot prompting with an LLM API
   • Compares on a held-out set using F1 per class as well as accuracy
   • Weighs cost, speed and data privacy

3. Walk me through how you would build a RAG system over 5,000 internal policy documents.

   What a good answer covers

   • Describes document parsing and chunking, with chunk size trade-offs
   • Explains creating embeddings and storing them in a vector database
   • Covers retrieval, optional re-ranking and prompt construction
   • Adds source citations in the answer
   • Describes how to evaluate retrieval and answer quality separately

4. When would you choose pgvector over a managed vector database such as Pinecone?

   What a good answer covers

   • Notes pgvector suits teams already running PostgreSQL and wanting one system
   • Mentions simpler data governance and joins with relational data
   • Notes managed services reduce operations work and scale more easily
   • Considers dataset size, query speed and where data may be hosted

5. Your fine-tuned model scores well in testing but gives poor answers in production. What do you check?

   What a good answer covers

   • Looks for differences between test data and live data
   • Checks for data leakage between training and test sets
   • Reviews preprocessing differences between training and serving
   • Samples real failures and labels them by type
   • Sets up monitoring for drift

6. How do you evaluate an LLM-based question-answering feature where there is no single correct answer?

   What a good answer covers

   • Builds a fixed test set from real user questions
   • Uses human review with a clear rubric
   • Uses automated checks for groundedness and citation accuracy
   • Tracks hallucination rate, latency and cost
   • Reruns the set after every prompt or model change

7. When would you fine-tune an open model from Hugging Face instead of calling a hosted LLM API?

   What a good answer covers

   • Cites data that cannot leave the organisation's environment
   • Cites high volume where per-call cost is a problem
   • Cites a narrow task where a small tuned model matches a large one
   • Acknowledges the extra work of training, hosting and maintenance

1. A business team wants a chatbot that answers staff questions about HR policy. They want it live in a month. How do you scope it?

   What a good answer covers

   • Clarifies the question types and the source documents
   • Proposes a small first release limited to well-covered topics
   • Plans a RAG design with citations and a fallback to a human
   • Defines success measures before building
   • Flags risks such as outdated documents

2. Your summarisation tool invents a clause that is not in the contract. A legal officer has noticed. What do you do?

   What a good answer covers

   • Acknowledges the issue and pauses use for high-risk cases
   • Reproduces the error and finds the cause, such as poor retrieval or prompt wording
   • Adds grounding checks and requires quotes from the source
   • Adds the case to the test set
   • Agrees a human review step for legal outputs

3. Your LLM API bill has doubled in a month. How do you find the cause and reduce it?

   What a good answer covers

   • Checks usage logs by feature, prompt length and call count
   • Looks for repeated calls, long contexts or retries
   • Considers caching, shorter prompts and smaller models for easy tasks
   • Tests that quality holds after changes
   • Sets usage alerts

1. Tell me about a model or project you built that did not work. What did you change?

   What a good answer covers

   • Describes a real project and the specific failure
   • Shows how the cause was found with data, not guesses
   • Explains what changed and the result
   • Shares what the person now does differently

2. Describe a time you explained a model's limits to a manager who wanted more than it could deliver.

   What a good answer covers

   • Gives a concrete situation
   • Uses everyday language and examples of errors
   • Offers a realistic alternative
   • Shows the outcome and how trust was kept

1. You are training a model on customer chat logs that contain names, NRIC numbers and phone numbers. How do you handle this under the PDPA?

   What a good answer covers

   • Checks the purpose and consent covered by the original collection
   • Removes or masks personal data before training where possible
   • Limits access and keeps data in approved storage
   • Considers data retention and the risk of the model repeating personal data
   • Involves the data protection officer

2. Many Singapore users write in Singlish or mix English with Mandarin, Malay or Tamil. How would you adapt an NLP system for this?

   What a good answer covers

   • Collects and labels real local text for testing
   • Checks whether the base model handles code-switching
   • Considers multilingual or locally tuned models
   • Reports accuracy by language group
   • Plans for slang and spelling variants

3. A bank wants an LLM assistant for relationship managers. What risk and governance points would you raise given Monetary Authority of Singapore (MAS) expectations?

   What a good answer covers

   • Knows MAS has issued guidelines on AI risk management for financial institutions
   • Proposes human oversight for customer-facing advice
   • Asks for testing, logging and explainability of outputs
   • Raises data security and vendor risk when using external LLM APIs
   • Suggests involving risk and compliance teams early

Common questions about a NLP engineer job

Sources