Neuronify
Hiring data as of September 20, 2026

AI skills in demanddefined, then counted

A glossary of the skills, tools, and job titles New York AI employers ask for. Each term gets a plain definition and a number: how many companies name it in open roles today.

Browse 60 terms
NYC AI companies read
47NYC AI companies read
open roles read today
1,567open roles read today
terms tracked
90terms tracked
terms with a full page
40terms with a full page

Why a glossary with numbers

AI vocabulary turns over faster than anyone can keep up with, and most glossaries answer only half the question. Knowing what evals or RAG means is useful. Knowing that one term is named by half of the city’s AI employers and another by three of them tells you which to learn, which to put in a job post, and which is already fading.

So every term here is measured against the same source: the open roles on the job boards of the companies in the NYC AI 100, read daily for the NYC AI Hiring Index. Definitions are written for two readers — the person deciding what to learn next, and the person writing the job description.

If you are buying AI rather than hiring for it, the companion AI glossary for buyers covers the terms that show up in vendor proposals and what each one costs you.

The most-named AI terms

Companies naming each term in at least one open role, out of 47. Programming languages, cloud platforms, and the catch-all “machine learning” are left out of this ranking and listed further down.

Three reads on today’s numbers

The market is for builders on models, not builders of them. 37 companies name AI agents, 25 name evals, and 20 name RAG. PyTorch is named by 11 and fine-tuning by 12. Both are real careers. They are different ones, and the second is the smaller market by a wide margin.

Nobody has won the model layer. Anthropic is named by 24 companies, OpenAI by 20, and Google Gemini by 9, and many postings list two or three together. Knowing how to compare models on a task is worth more than knowing one provider’s API.

Retrieval runs on what companies already had. Elasticsearch is named by 9 companies. Pinecone, the best-known dedicated vector database, is named by 2. Search engineering experience transfers more directly than the tooling landscape suggests.

All 60 terms

Grouped by what kind of work they describe, most-named first within each group. Linked terms have a full page with the companies hiring, pay, role mix, trend, and example open roles. For scale: the median posted base across all New York engineering roles is $215k.

Building with LLMs

The application layer: what most AI engineering jobs in New York actually consist of.

TermCompaniesOpen rolesMedian pay
LLMs (large language models)Models trained on very large text corpora to predict the next token, which turns out to be enough to write, summarize, extract, translate, and reason over language.40261$213k
AI agents (agentic AI)Systems where a language model decides which steps to take — calling tools, reading results, and continuing — instead of answering once and stopping.37261$220k
Evals (LLM evaluation)Repeatable tests that score a model or agent’s outputs against examples with known good answers, so a change can be judged by measurement instead of by feel.2549$230k
RAG (retrieval-augmented generation)Fetching relevant documents at question time and putting them in the model’s context, so answers come from your data instead of the model’s memory.2031$238k
Generative AIAI systems that produce new content — text, images, audio, video, code — rather than only classifying or scoring existing data.17103$183k
Prompt engineeringDesigning and testing the instructions and examples given to a language model so it performs a task reliably.1626$175k
Tool use (function calling)Letting a language model call external functions and APIs — search, databases, code execution — by emitting structured requests the application then runs.1013—
Multimodal AIModels that take in or produce more than one kind of data — text, images, audio, video — within a single system.713—
GuardrailsChecks placed around a model’s inputs and outputs to block unsafe, off-policy, or malformed results before they reach a user or a downstream system.611—
EmbeddingsNumeric vectors that represent the meaning of text, images, or other data, so that similar items sit close together and can be found by distance.56—
Context engineeringDesigning everything a model sees at inference time — instructions, retrieved documents, tool results, memory — as one budgeted system.44—

Model providers and frameworks

Named vendors and libraries. A provider in a posting means a choice already made; a framework is more often shorthand for experience.

TermCompaniesOpen rolesMedian pay
Anthropic (Claude)Maker of the Claude family of language models and the Claude Code coding agent. One of the two providers New York AI companies name most.2464$208k
OpenAI (GPT)Maker of the GPT models and ChatGPT. One of the two model providers New York AI companies name most in job posts.2037$173k
MCP (Model Context Protocol)An open protocol for connecting AI applications to tools and data sources through a standard interface, instead of a custom integration for each pairing.1328—
Elasticsearch (OpenSearch)A distributed search engine for keyword, and now vector, search. The most-named retrieval technology in New York AI postings.931$224k
Google Gemini (Vertex AI)Google’s family of multimodal models, reached through the Gemini API or Vertex AI on Google Cloud.920$251k
LangChainAn open-source framework for building applications on language models — chaining calls, connecting retrieval, and managing tools and memory.921—
LangGraphA framework from the LangChain team for building agents as explicit graphs of steps, with state, branching, and human approval points.66—

Model development

Training, adapting, and shrinking models. A smaller market than the application layer, with its own tools and its own interview loop.

TermCompaniesOpen rolesMedian pay
Machine learningBuilding systems that learn patterns from data instead of following hand-written rules. The parent field of everything else in this glossary.41197$215k
Fine-tuningContinuing to train an existing model on your own examples so it adopts a format, tone, or narrow skill that prompting alone does not reliably produce.1222—
PyTorchThe dominant open-source framework for building and training neural networks, and the clearest marker of a role that works on models directly.1133$265k
TensorFlowGoogle’s deep learning framework. Once the default; now mostly named alongside PyTorch or in teams with older production models.619—
Hugging FaceThe hub where open models and datasets are shared, and the Transformers library most teams use to load and fine-tune them. Headquartered in New York.518—
Distributed trainingTraining one model across many GPUs or machines at once, which brings its own engineering problems in communication, checkpointing, and failure recovery.46—
Quantization and distillationTwo ways to make a model cheaper to run: store its weights at lower precision, or train a smaller model to imitate a larger one.44—
Reinforcement learning (RLHF)Training a model by rewarding good outcomes rather than showing correct answers; RLHF applies it to language models using human preference ratings.311—
Deep learningMachine learning with many-layered neural networks. The term has largely dropped out of postings, replaced by the specific model families built on it.22—
Synthetic dataTraining or evaluation data generated by a model or a simulator, used where real examples are scarce, private, or expensive to label.23—

Infrastructure and operations

What keeps AI systems running and affordable once they are in front of users.

TermCompaniesOpen rolesMedian pay
ObservabilityInstrumenting a system so its internal state can be understood from its outputs — logs, metrics, traces — and, for AI systems, the full record of each model call.35128$225k
Data pipelines (ETL)The jobs that move data from where it is created to where it is used, cleaning and reshaping it on the way. Every AI system sits on top of one.3073$213k
InferenceRunning a trained model to produce outputs. For deployed AI products it is where most of the compute bill, and most of the latency, comes from.1437$238k
MLOps (ML platform)The engineering that gets models into production and keeps them there: deployment, versioning, monitoring, and the platform other teams build on.1321$215k
GPUsThe parallel processors that train and serve modern models. In a posting, the word marks a role that works close to the hardware.519$238k

Data platform

Warehouses, orchestration, and streaming. These mark the data and analytics roles inside AI companies.

TermCompaniesOpen rolesMedian pay
dbtA tool for defining data transformations as version-controlled SQL, with testing and documentation. The backbone of the analytics engineering role.1737$185k
SnowflakeA cloud data warehouse. In AI company postings it marks the analytics and data-platform side of the house.1747$222k
Apache KafkaA distributed event-streaming platform for moving large volumes of data between systems in real time.1038$231k
Apache AirflowThe most widely used open-source scheduler for data pipelines, which are defined in Python as graphs of dependent tasks.923$235k
BigQueryGoogle Cloud’s serverless data warehouse, queried with SQL.914$215k
DatabricksA data and AI platform built around Apache Spark and the lakehouse architecture, used for large-scale data engineering and model training.68—
DagsterA newer data orchestrator built around the data assets a pipeline produces rather than the tasks it runs.59—

Applied domains

Problem areas with their own methods. Several are concentrated in one or two New York employers.

TermCompaniesOpen rolesMedian pay
NLP (natural language processing)The field concerned with getting computers to work with human language. LLMs absorbed most of it, and the term is fading from postings.517—
Recommender systemsModels that predict what a user will want next, from products to content to matches on a marketplace.44—
Computer visionModels that interpret or generate images and video — in New York, mostly at generative-media and robotics companies.35—
Knowledge graphsStructured maps of entities and the relationships between them, increasingly used to give retrieval systems something firmer than text similarity.24—
Search and rankingOrdering results by relevance: classical information retrieval, learning-to-rank models, and the offline metrics used to judge them.24—
Voice AISpeech recognition, speech synthesis, and the voice agents built from them. Heavily concentrated in a few New York employers.267—

Safety and governance

Thinly represented in startup postings, and a much larger market among the city’s banks, insurers, and law firms.

TermCompaniesOpen rolesMedian pay
Human-in-the-loopDesigning a workflow so a person reviews, corrects, or approves what the model produces at the points where a mistake would matter.919—
AI safety (responsible AI)Work to keep AI systems from causing harm: abuse prevention, red-teaming, safety evaluations, and the policies that decide what a product will not do.710—
AI governanceThe policies, controls, and accountability an organization puts around how AI is built, bought, and used — and the evidence that they are followed.621—

Emerging job titles

Matched on the job title rather than the description. New specialties show up in descriptions first and in titles later.

TermCompaniesOpen rolesMedian pay
Forward deployed engineerfull guideAn engineer embedded with customers to get the product working inside their systems. The most common emerging title in our data, and the one with the fastest-growing search interest.1345$220k
AI engineer (applied AI)full guideA software engineer who builds products on top of existing models — the application layer, as distinct from training models.710—
Solutions architectA technical pre-sales or post-sales role that designs how a product fits a customer’s systems and proves it will work before they buy.69—
Developer relationsDeveloper advocates and developer-relations roles, which AI infrastructure companies hire to win engineers over.23—
Agent engineerAn engineer whose title names agents specifically. Still rare as a title, even though agent work is named in a large share of descriptions.11—
AI safety (in the job title)Roles titled for AI safety, alignment, or trust and safety.11—
Evals (in the job title)Roles with evaluation in the title itself. Far rarer than descriptions that ask for evals, which is how new specialties usually start.11—
Context engineerA title we track for its first appearances. The practice is being named in descriptions well before it shows up in titles.00—
ML infrastructure engineerAn engineer who builds the platform models are trained and served on, identified here by “ML infrastructure” or “ML platform” in the title.00—
Prompt engineerThe 2023 job title. We track it to document its decline; the skill moved into other roles.00—

Also tracked

Languages and cloud platforms are our control group: if Python stopped appearing in most engineering postings, the extraction would be broken, not the market. The rest are tools named by too few companies so far to say anything about. They earn a definition when that changes.

ToolCategoryCompaniesOpen rolesMedian pay
PythonLanguages42328$215k
AWSCompute39201$235k
TypeScriptLanguages39151$222k
SQLLanguages31168$180k
GoLanguages29110$235k
KubernetesCompute27119$222k
GCPCompute2097$222k
AzureCompute1756$241k
JavaLanguages1669$243k
KotlinLanguages918$215k
RustLanguages921$255k
CUDALanguages45—
JAXTraining frameworks34—
LangfuseEvals & observability34—
RubyLanguages33—
SwiftLanguages33—
ArizeEvals & observability24—
C#Languages25—
PineconeVector & retrieval25—
SGLangServing & inference24—
SparkData platform24—
TensorRTServing & inference22—
TritonServing & inference22—
vLLMServing & inference24—
AutoGenOrchestration & agents16—
BasetenServing & inference11—
BraintrustEvals & observability11—
CrewAIOrchestration & agents15—
LlamaIndexOrchestration & agents11—
Meta LlamaModel providers11—
ModalServing & inference11—
RayServing & inference11—
ScalaLanguages12—

How the numbers are made

Source. Once a day we read every open role on the public job boards of the New York AI companies we cover, through the posting APIs that Greenhouse, Ashby, and Lever provide for exactly this purpose. Descriptions vanish when a role is filled, so tags are extracted the same day and the text is discarded.

Companies first. The headline figure for every term is the number of companies naming it, not the number of postings. A single employer can have two hundred roles open; counted by posting, its vocabulary would become the market’s.

Boilerplate removed. Most companies open every posting with the same paragraph about themselves. Before matching, we drop any sentence a company repeats across its postings, so a voice-AI company’s finance roles are not counted as voice-AI jobs. When we tested this on 1,131 postings it left company counts almost untouched and cut some posting counts by more than four-fifths.

Ordinary words handled strictly. Sales postings “fine-tune” messaging and finance postings set “guardrails”. Terms like these count only beside the model or agent language that gives them their technical meaning, acronyms are matched case-sensitively, and where a term cannot be separated from its everyday sense we leave it out. A vendor’s own job posts never count toward its own name.

Pay. New York roles only, from the salary ranges employers post under the city’s pay transparency law. Base salary only. A range repeated across one company’s postings counts once, and no median is shown below eight distinct ranges. A term’s pay figure describes the roles that name it; it is not a premium for the skill.

The full method. Everything above, plus the limits of the coverage and a dated log of changes, is on the methodology page.

What it does not say. Named in a posting is not used in production, and a nice-to-have counts the same as a requirement. Our coverage is startup-weighted, so terms that matter most inside banks, insurers, and law firms — governance above all — are understated.

AI skills glossary — common questions

What AI skills are most in demand in New York?

Measured by how many companies name them in open roles on September 20, 2026: LLMs (40 of 47 companies), AI agents (37 of 47 companies), Observability (35 of 47 companies), Data pipelines (30 of 47 companies), Evals (25 of 47 companies), Anthropic (24 of 47 companies). The pattern is consistent: employers ask for the ability to build and evaluate applications on existing models far more often than for the ability to train models.

How is this different from other AI glossaries?

A standard glossary defines terms. This one also measures them: every entry carries the number of New York AI companies naming it in open roles today, the posted pay on those roles where there is enough data, and the other terms it appears beside. The definitions are written for someone reading a job post or writing one, not for an exam.

Where does the data come from?

From the public job boards of the companies on the NYC AI 100 and the AI in NYC Show roster, read once a day through the Greenhouse, Ashby, and Lever posting APIs. We extract the tools and concepts each description names, keep only those tags, and discard the text. Nothing is scraped from aggregators and no figure is estimated.

Does a term being named mean the company uses that technology?

No. It means a job posting named it, as a requirement or as a nice-to-have. That is a measure of what employers ask candidates for, which is usually close to what they run, but it is not an audit of anyone’s production stack and we never describe it as one.

Why do some terms have no pay figure?

We publish a median only when at least eight distinct posted salary ranges exist on New York roles naming the term. Below that, a median describes two or three employers rather than a market. Pay is base salary from employer-posted ranges, which New York’s pay transparency law requires; commission-based on-target earnings are excluded.

How often is the glossary updated?

The numbers refresh daily. Definitions are revised as usage shifts, and a new term is added when it starts appearing across several companies’ postings — which is itself worth noticing, since a term entering the vocabulary of job posts is an early sign of a skill becoming a hiring category.

Hiring for one of these?

A term in a job post is easy to match and hard to verify. We match companies with AI engineers and consultancies vetted on production work — and if you have the skills on this page, we would like to know you.

Start a project briefJoin the expert network