Each section below stands on its own. If you are trying to decide which requisition to open, read the "hired to" line first and ignore the title.
Applied AI / LLM Product Engineer
What AI engineers are startups hiring?
Hired to: Turn a foundation model into a feature customers pay for. These engineers usually do not train models. They are accountable for whether the AI part of the product works in front of a user.
What they own: The AI surface of the product. Retrieval, prompting, tool calling, structured output, latency, and the evaluation harness that tells the team whether a change made the feature better or worse.
Signals employers screen for
- Has shipped an AI feature to production users, not a demo or notebook
- Can describe a specific evaluation they built and what it caught
- Talks about failure modes and fallbacks before talking about model choice
- Has an opinion on retrieval quality grounded in something they measured
Why it is hard to fill: The title fragmented. Companies now separate AI Product Engineer, ML Platform Engineer, LLM Engineer, and AI Research Engineer with different compensation curves, so identically titled candidates are frequently not comparable.
Applied AI and LLM engineer searches grew 47 percent year over year in Recruits Lab's 2025 search data — the largest engineering category increase in our practice. 2026 AI Hiring Report
Agentic AI Engineer
What is an agentic AI engineer?
Hired to: Take an agent from an impressive demo to something that can be given real permissions against real systems. The job is mostly reliability engineering wearing an AI title.
What they own: Systems that plan and act rather than answer. Tool permissioning, memory, retries and rollback, guardrails, and the evaluation work that makes an agent trustworthy enough to run without a human confirming every step.
Signals employers screen for
- Has run an agent in production with real side effects, not read-only tasks
- Can describe the permission model and what they refused to let the agent do
- Has built evaluations for multi-step behavior, not single-response quality
- Talks about cost and loop control as first-class design problems
Why it is hard to fill: Role definition is inconsistent across companies, and the pool with genuine production agent experience is small. Many candidates have built agent prototypes; far fewer have operated one under load with real permissions.
Practitioner observation from active Recruits Lab searches: agentic scope is most often added to an existing applied AI requisition rather than opened as its own headcount, which is a common source of mis-scoped searches.
Machine Learning Engineer
What is the difference between an ML engineer and an AI engineer?
Hired to: Fix or own a measurable model-quality bottleneck. Companies hire here when the problem is the model, not the interface to it.
What they own: Model quality. Data and feature pipelines, training decisions, serving infrastructure, and the measured performance of the model itself rather than the product surface around it.
Signals employers screen for
- Owns a metric, and can say what it was before and after their work
- Comfortable with data quality problems that look unglamorous
- Has made a build-versus-fine-tune-versus-API call and can defend it
Why it is hard to fill: Demand shifted toward platform and serving work faster than the candidate pool re-labeled itself, so strong ML platform engineers are often hidden behind generic ML titles.
ML platform engineer demand grew 33 percent year over year in Recruits Lab's 2025 data, driven by serving infrastructure and evaluation tooling needs. 2026 AI Hiring Report
AI Research Engineer
Are companies hiring AI research engineers?
Hired to: Improve the model itself, or adapt a foundation model in ways an API cannot. At non-frontier companies this usually means post-training and evaluation rather than pretraining from scratch.
What they own: Moving research into working systems. Pretraining and post-training, reinforcement learning from human feedback, multimodal and speech work, evaluation research, and the experimental infrastructure that makes results reproducible.
Signals employers screen for
- Has a reproducible result, published or internal, that someone else relied on
- Can explain a negative result and what it ruled out
- Distinguishes clearly between pretraining, post-training, and evaluation work
Why it is hard to fill: This is the slowest search category in our practice and the most mandate-sensitive. Candidates frequently choose research direction and compute access over cash.
AI research engineer searches at non-frontier startups grew 28 percent year over year in Recruits Lab's 2025 data, with talent moving from academia and large platform companies. Time-to-fill averages 12 to 20 weeks. 2026 AI Hiring Report
AI Infrastructure Engineer
What AI infrastructure roles are in demand?
Hired to: Make training and inference fast enough and cheap enough to keep shipping. Companies usually open this role after scale or spend becomes the constraint, not before.
What they own: The compute layer. Distributed training, inference serving, GPU platform operations, throughput and latency, and the cost curve underneath every AI feature the company ships.
Signals employers screen for
- Has reduced a real inference cost or latency number and can quantify it
- Has operated multi-node training, including the failure recovery
- Understands the hardware well enough to explain where the time actually goes
Why it is hard to fill: The skill set sits between systems engineering and ML, and the strongest people are usually employed at companies where compute is the core product.
Practitioner observation from active Recruits Lab searches: infrastructure requisitions tend to follow a cost or reliability event rather than a hiring plan, which compresses the timeline available to fill them.
Forward Deployed Engineer
What is a forward deployed engineer?
Hired to: Get enterprise customers to production when the product alone will not do it. Companies open this role when deals stall at implementation rather than at the demo.
What they own: Deployment of the product inside a customer's environment, and the integration code required to make it work there. An FDE writes software, sits with the customer, and carries responsibility for the outcome on the customer's side.
Signals employers screen for
- Has written and shipped code inside a customer's stack, not just configured it
- Comfortable being the technical face of the company in a customer room
- Can describe an integration constraint they discovered on site and solved
Why it is hard to fill: It requires two profiles that rarely coexist: genuine engineering depth and customer-facing composure. Candidates who are strong at one are usually weak at the other, and the title means different things at different companies.
Practitioner observation from active Recruits Lab searches: FDE searches most often fail at the scoping stage, because the company has not decided whether the hire is an engineer who talks to customers or a solutions consultant who can code.