NLP Engineer Job Description Template 2026

Last updated: September 15, 2026

By Tom Kenaley, President and Senior Partner, KORE1

An NLP engineer job description works when it names the text the hire will process, the languages and daily volume, whether labeled training data exists, and whether the seat fine-tunes its own models or builds around LLM APIs. Pay? Figure $130,000 to $175,000 base for mid-level in 2026, and $180,000 to $240,000 once the person is senior enough to own a production system. Our NLP engineer staffing desk runs into the same few scoping gaps on nearly every req we see, and the template below is built around them.

Eleven weeks. That’s how long a property and casualty insurer in Columbus had been running its NLP engineer req when the claims VP forwarded it to us.

The title was “NLP Engineer (LLM / GenAI).” Underneath sat transformers, LangChain, RAG, fine-tuning, chatbots, sentiment analysis, spaCy, NLTK, “PhD preferred,” and a $160,000 base. Forty-one people applied. Most had built chatbot demos. Two had shipped something to production that read insurance documents, and neither got past the first screen, because the screen was checking for LangChain.

So what did the job actually need? About 1.8 million adjuster notes and claimant emails came in every month. The claims team wanted eleven fields out of each one (date of loss, any mention of an injury, a hint that some third party might owe subrogation, that kind of thing) so the fraud review queue could sort itself, instead of four people with highlighters doing it badly by Thursday. Someone had prototyped the extraction against a hosted LLM. It worked fine. Then the pilot bill for a single month of backfill landed at roughly $41,000. Nobody had labeled a single note by hand, either, so nobody could say whether the prototype was right 70% of the time or 95%.

We rewrote the posting around that. Eleven fields, the monthly volume, and a line saying the hire would write the first annotation guideline and fine-tune a smaller model to take over most of the API traffic. Same title, minus the parenthetical. The base moved to $165,000 to $185,000, because the person who can do all of that is senior, and pretending otherwise was part of why the first eleven weeks went nowhere. Their hire accepted 26 days after the repost.

Worth saying plainly, since you’re reading a staffing firm’s blog. We’re paid by the employer on placements, so a posting clear enough to fill the seat on its own costs us the search sometimes. Fine by me. Write it anyway.

What the Text Looks Like

An NLP engineer builds and maintains software that reads human language at scale, turning emails, clinical notes, contracts, chat logs, and search queries into structured fields, labels, rankings, or redactions other systems can use. In 2026 that means training and evaluating task-specific models about as often as it means calling a large language model.

Three facts about your text sort the applicant pool faster than any skills list. What it is. How much of it shows up, and how fast an answer has to come back. And whether anyone has labeled a sample by hand yet.

The text problem itself usually falls into one of these.

Most real seats are two rows. Extraction plus redaction is common in healthcare. Routing plus search shows up all over customer support, where a ticket classifier and a help-center search engine end up sharing an embedding model and, eventually, one very tired engineer.

Put the languages in, and be specific about them. “English and Spanish” undersells the job when your Spanish arrives as WhatsApp messages from Puerto Rico, half Spanglish, abbreviated to the bone, because a model trained on clean news text will miss a lot of it. Say where the text comes from. Candidates who have handled informal or code-switched language will spot that sentence and apply because of it.

NLP Engineer Responsibilities Worth Writing Down

A lot of templates online promise the hire will “develop innovative NLP solutions to drive business value.” Which solutions? For what text? Senior people stop reading right about there.

On the NLP searches we fill, the week tends to look more like this.

• Reading a few hundred real documents before training anything. Actually reading them, with someone from the business, because the text is always stranger than the ticket described.
• Annotation guidelines, written and then rewritten, plus checks on whether two labelers agree.
• Fine-tuning encoders like ModernBERT or DeBERTa in Hugging Face Transformers, or training a spaCy pipeline when rules and a small model will do.
• Evaluation. Precision, recall, and F1 per field and per document type, on a held-out set nobody tuned against.
• LLM calls where they earn their cost, usually for the long tail of ugly documents a smaller model can’t handle.
• Shipping the model behind an API or a nightly batch job, then tracking latency and cost per thousand documents week over week.
• A vendor changes an invoice template, or a new product brings in words the classifier has never seen. Somebody has to notice the drift before the claims team does.
• Explaining to a compliance officer why the model got one specific document wrong, and what fixing it would take.

No chatbots on that list. Some NLP seats build conversational systems, and if yours does, say so near the top. Most don’t. Putting “chatbot experience” on a document extraction req attracts exactly the wrong pile of resumes, and then you spend three weeks politely declining them.

A Smaller Model or a Hosted LLM

This is the sentence I most want to see in NLP postings. I almost never do.

A hosted LLM can extract fields, classify tickets, and tag entities with a decent prompt and zero training data. For a pilot, or for a few thousand documents a day, that’s often the right call, and a good NLP engineer will tell you so without being asked. Volume changes the answer. At millions of documents a month, per-token pricing starts to look like a second salary, latency creeps up, and every customer email is leaving your network for a third party, which your security team probably has opinions about.

The alternative got a lot better recently. ModernBERT, from Answer.AI and LightOn, is an encoder that reads up to 8,192 tokens where most older BERT-style models stopped at 512, and its base version has only 149 million parameters (the ACL 2025 paper behind it has the benchmark results). Fine-tuned on a few thousand labeled examples, a model that size runs on modest GPUs and, on a narrow extraction task, can hold its own against something far larger. It also stays in-house. Hugging Face, for its part, shipped Transformers v5, the library’s first major version in five years, which went PyTorch-only and rebuilt tokenization, so “Transformers experience” on a 2023 resume describes a somewhat different library than the one your team may be running now.

You don’t need to choose the approach in the posting. You do need to hand candidates the numbers they’d use to choose. Two lines cover it.

• “We process about [N] documents a [day / month], and results need to come back within [latency].”
• “Today this runs on [a hosted LLM / rules / a vendor tool / people], and we want [lower cost / better accuracy / text kept in-house].”

An engineer who has done that migration before reads those lines and starts sketching an architecture in their head. That’s who you want applying.

If the seat is mostly prompts, model selection, and serving a large model, you’re hiring a different role, and our LLM engineer job description is the better place to start. General model training with no language focus fits the machine learning engineer job description template instead. Product features built on a foundation model fit the generative AI engineer job description better, and if serving and pipelines are the whole job, the MLOps engineer version is closer. For the umbrella title, there’s our AI engineer job description.

Somebody Has to Label the Data

Every production NLP system sits on top of a few thousand examples somebody labeled by hand. Postings almost never mention it. Experienced candidates notice.

A revenue cycle management company in Nashville learned this slowly in 2025. They hired a sharp NLP engineer out of a university research group, someone whose models had always trained on public benchmark datasets that showed up already labeled. The task was flagging which diagnoses in physician notes were current and which were history. Two contract medical coders labeled the training data. Three months in, F1 was stuck around 0.74, and nobody could explain it.

Nobody had checked whether the coders agreed with each other. Not once. When someone finally scored 400 notes that both had labeled, Cohen’s kappa came back at 0.58. One coder treated “history of type 2 diabetes, controlled on metformin” as a current condition. The other didn’t. Both readings were defensible, honestly, and the model was trying to learn two rules at once.

A nine-page guideline, one adjudication round, and five weeks later, agreement was at 0.83 and F1 at 0.88. Same engineer. Same architecture. What changed was a chunk of work the job description never mentioned, so the hire hadn’t expected to own it, and nobody else owned it either.

Say it in the posting. Does a labeled set exist? Who labels? In-house experts, contractors, a vendor? Which tool? Label Studio, Prodigy, Argilla, or a shared spreadsheet you’re a little embarrassed about? And is writing the guideline part of this job? For most first NLP hires it should be.

Clinical Notes, Contracts, and Text With Rules Attached

Some text shows up with legal strings attached. The posting is the cheapest place to mention them.

Clinical notes are the obvious case. HHS guidance on de-identifying protected health information lays out two routes under the HIPAA Privacy Rule, Expert Determination and Safe Harbor, and Safe Harbor means removing 18 kinds of identifiers. Structured fields are easy. Free text is harder. Picture a note that says the patient’s daughter drove her in from Tulsa after her shift at the elementary school. Relative. Place. Employer. Three identifiers in a line nobody would look at twice, and the pipeline that has to catch them ends up, more often than not, being the NLP engineer’s first real project in healthcare regardless of what the posting promised.

Other text has rules too. Contracts raise privilege questions. Support call transcripts carry card numbers people read aloud to agents, and HR text, like exit interviews and complaint logs, is sensitive in ways that don’t fit a neat regulation.

None of this should scare off a good candidate. It changes who fits, though. Anybody with two years inside a HIPAA environment already knows you don’t just pull a sample into a notebook. Give that person one line, something like “You’ll work with de-identified clinical notes inside [environment], and raw PHI never leaves [system],” and they’ll know the job on sight. Healthcare NLP pools run thin, which is one reason our healthcare IT staffing recruiters usually get pulled into these searches.

NLP Engineer Salary in 2026

There’s no federal job code for NLP engineers. The nearest one, computer and information research scientists, had a median wage of $140,300 in May 2025 according to the Bureau of Labor Statistics, and BLS expects employment in that group to grow 22% between 2025 and 2035, with roughly 2,900 openings each year. A decent anchor. That category also sweeps in academic researchers, so treat it as a midpoint rather than a ceiling.

The aggregators don’t agree. Glassdoor puts the average U.S. NLP engineer salary near $165,000, with the middle half of reports between about $138,000 and $200,000. ZipRecruiter showed an average of about $107,000 in September 2026, with most pay between $88,500 and $136,000. Almost sixty grand apart. My read on why? Plenty of postings with “NLP” in the title are really analyst or annotation jobs, and those pull a posting-based average down, while Glassdoor’s self-reported numbers lean toward tech companies that pay engineers well. Both are measuring something real, and neither one is your seat.

Those bands line up with our guide to hiring NLP engineers, which also covers what AI labs pay on top of base. Clinical, legal, and multilingual seats tend to land in the upper half of each range, since fewer people have done that work in production. Our salary benchmark assistant will price a particular city and stack for free.

Post the range. The Massachusetts pay range rule kicked in on October 29, 2025, and it covers any employer with at least 25 people working in the state. Colorado, Washington, and New York got there years earlier. NLP candidates also compare notes with each other more than you’d think. A blank where the number should go reads as a low number.

NLP Engineer Job Description Template

Replace everything in square brackets. The notes in parentheses are for you, not for candidates, so cut them before the posting goes up. Section doesn’t fit your team? Drop it.

Job Title

[NLP Engineer, Document Extraction / Senior NLP Engineer, Clinical Text / NLP Engineer, Search and Retrieval / Multilingual NLP Engineer] (Put the text problem in the title. “NLP Engineer” by itself draws chatbot builders, data analysts, and research scientists in roughly equal numbers.)

About the Role

[Company] is a [size] [industry] company that handles about [N] [documents / tickets / notes / queries] a [day / month]. We’re hiring an NLP engineer to [pull (fields) out of (document type) / route (ticket type) to the right team / improve search across (collection) / de-identify (text type) before it reaches analytics]. You’ll report to [head of data science / director of machine learning / VP of engineering] and work closely with [claims / clinical operations / support / legal] on what a correct answer looks like for each document.

The Text and the Data

• Sources: [adjuster notes / physician notes / contracts / support chats / product reviews], mostly in [English / English and Spanish / (languages)], and [clean and templated / informal and full of abbreviations].
• Volume and speed: roughly [N] a [day / month], with results needed [in real time, under (N) milliseconds / in a nightly batch].
• Labeled data: [a gold set of (N) examples exists / a partial set exists with no written guideline / nothing yet, and you’ll build it]. Labeling is done by [in-house experts / contractors / a vendor] in [Label Studio / Prodigy / Argilla / spreadsheets].
• Current approach: [a hosted LLM / rules and regular expressions / a vendor product / manual review], and we want to improve [cost / accuracy / latency / keeping text in-house].
• Restrictions: [HIPAA / PCI DSS / attorney-client privilege / none], and text [stays inside (environment) / can be sampled locally after de-identification].

What You’ll Do

• Read real samples of our text with [team] and define what the model should produce for each document type
• Write the annotation guideline, check agreement between labelers, and resolve disagreements before training starts
• Build and fine-tune models for [extraction / classification / retrieval / redaction] using [Python, PyTorch, Hugging Face Transformers, spaCy]
• Decide, with numbers, where a hosted LLM belongs in the pipeline and where a smaller model does the job for less
• Report precision, recall, and F1 by [field / class / language] on a held-out test set, and keep that set clean
• Deploy models to [our API / a batch pipeline on (platform)] and track latency and cost per thousand documents
• Catch drift when [forms / vendors / products / patient populations] change, and retrain on a schedule we agree on
• [Regulated seats: maintain the de-identification or redaction step and the audit sampling that shows it works]

What You Bring

• [3+ / 6+] years building NLP systems, including at least [one / two] that ran in production on real, messy text
• Strong Python and hands-on PyTorch, with fine-tuning experience on encoder models such as BERT, RoBERTa, DeBERTa, or ModernBERT
• Experience creating labeled data, not only using it, ideally including a guideline you wrote or revised
• Evaluation habits that go past a single accuracy number, like per-class metrics, error analysis, and a test set kept away from tuning
• [Retrieval seats: hybrid search with Elasticsearch or OpenSearch plus embeddings, and experience collecting relevance judgments]

Nice to Have

• Production work with hosted LLM APIs, including cost tracking and a fallback for when the API is slow or down
• Text experience in [insurance claims / clinical documentation / contracts / customer support]
• [Language] fluency or earlier multilingual NLP work (List this only if the model really has to handle that language. Otherwise it quietly shrinks the pool.)
• [Regulated seats: prior work inside a HIPAA, PCI DSS, or similarly controlled environment]

How We Decide a Model Is Ready

[Two or three sentences. Which metric, on which test set, has to reach what number before a model replaces the current approach, and who signs off. For example, “Field-level F1 of 0.90 or better on a 1,000-document held-out set, reviewed by our claims operations lead.”] (Candidates with real production experience read this part closely. No bar yet? Say the hire will help set it. That’s honest, and good engineers tend to like being handed that job.)

Pay, Location, and Schedule

Base: [$X to $Y]. On top of that, [bonus / equity] and [benefits]. [Remote / hybrid with (N) days in (city) / on site in (city)]. On-call for production models: [yes, on a rotation of (N) / no]. (Use a real range. Clinical, legal, and multilingual seats justify the top half of it.)

NLP Hiring, Answered Briefly

What does an NLP engineer actually spend the week building?

Pipelines that turn raw language into something another system can act on, such as extracted fields, routed tickets, ranked search results, or redacted records, along with the labeled data and tests behind them.

The mix shifts by team. An extraction seat spends a lot of time on document formats and edge cases, and a search seat mostly lives inside relevance judgments. Almost every one spends more time on data than the engineer expected in year one.

NLP engineer or LLM engineer, does the title really matter?

It matters once you know your volume and whether you own labeled data, because NLP engineers typically train and evaluate task models, while LLM engineers work mostly with prompts, model selection, and serving around large models.

Overlap is real. Plenty of people do both. Pick the title that matches the first six months of work, and if those months are mostly RAG pipelines and agent behavior, the LLM title will draw a better pool.

Do we actually need a PhD for this?

Rarely, unless the team publishes research or designs new model architectures, since most production NLP work rewards shipped systems and careful evaluation more than a dissertation does.

A PhD requirement also cuts the pool hard. Keep it under preferred if you like the signal. Personally I’d rather see “wrote the annotation guideline” on a resume than a paper on a benchmark nobody uses in production, but that’s a recruiter talking.

How much does an NLP engineer make in 2026?

$130,000 to $175,000 base, for a mid-level NLP engineer in the United States this year, while a senior engineer who owns a production system typically lands somewhere between $180,000 and $240,000.

Frontier AI labs pay well above that once equity is counted. Clinical or multilingual depth pushes an offer toward the top of the band.

Which libraries and tools should the posting name?

The ones already running in your stack, which for most teams in 2026 means Python, PyTorch, Hugging Face Transformers, and often spaCy, plus whatever search engine or labeling tool the work depends on.

Leave NLTK off unless it’s genuinely in production. Most working engineers treat it as a teaching and prototyping library these days, so seeing it on a req suggests the posting was copied from an old template. Candidates pick up on that.

Realistically, how long does an NLP search take?

Six to ten weeks is common for a senior NLP engineer, longer than the 17-day average KORE1 sees across IT searches, because relatively few people have shipped language models into production.

The fastest NLP searches we run have a clearly named text problem and a real range from day one. The slowest start out as “NLP Engineer (LLM / GenAI)” and get rewritten in week eight.

After the Req Is Written

A posting built this way does one job well. It lets the right engineer recognize their own work before they reach the requirements. Screening is a separate problem, and our NLP engineer hiring guide covers how to test for real production experience, where these searches stall, and when an NLP hire isn’t the answer at all.

Pilots and migrations, like moving an extraction job off a hosted LLM, often suit contract staffing for a defined stretch. The engineer who owns the model for years is a direct hire. Usually. Searches that straddle NLP, LLM, and classical ML work run through our AI and ML engineer staffing practice. If your req has been open a while, or you’re about to write your first one, talk to a recruiter on our AI team and we’ll tell you straight whether it reads like the job you have. A year in, 92% of KORE1 placements are still sitting in the seat they were hired for.