Hire LLM engineers who have actually shipped production models.

The title has drifted. Three years ago an LLM engineer was someone running `OPENAI_API_KEY=... python agent.py` in a notebook. Today, for a Series B product scale-up, it's the person who owns the end-to-end cost, quality, and latency of the LLM-backed product surface — usually three separate jobs at once: prompt design and output evaluation, retrieval and context shaping, and the serving-side optimisation that turns a 900ms p95 into a 180ms one. Getting all three to coexist inside a single engineer's head is rare. Getting them to coexist alongside product-engineering fluency (tests, deploys, observability, PR discipline) is rarer still, and that rarity is exactly what's distorting UK and Western European senior-AI comp bands.

Demand reflects that. According to the 2025 Stack Overflow Developer Survey, 62% of respondents worked with AI tools in production, but only 11% reported shipping an LLM-backed feature that served real users at sustained volume. That gap — between tinkering and operating — is where the 2026 hiring bottleneck sits. Every series-B company we have talked to in the last six months has at least one open LLM engineering role, and most of them have been open for more than 90 days. The generic cure is "hire three, accept one" — but that ties up 40–60 hours of senior engineer interview time per placement, which is the most expensive part of the whole exercise.

The roles fall into three archetypes we see most often: (1) the LLM product engineer, who owns a single AI feature end-to-end and is 60% application engineer, 40% ML; (2) the LLM platform engineer, who builds the internal inference/eval infra that feature teams consume; (3) the LLM research-leaning engineer, who moves between fine-tuning, distillation and RL workflows. Most scale-ups need archetype 1 first and archetype 2 at ~30 engineers; archetype 3 is almost always an over-hire below Series C. If you are unsure which archetype you actually need, the test is simple: would a 3-person team of archetype-1 engineers ship faster than a single archetype-3 engineer would? Below ~15 AI engineers the answer is almost always yes.

The 2026 market shift that is worth calling out: the gap between a strong LLM product engineer and a weak one has widened, not narrowed, since open models caught up in quality. When every team has access to Claude 4.7, Llama 3.3 70B, Qwen 3, and GPT-5, the differentiator is no longer model access — it is how well the engineer shapes context, evaluates output, and operates inference at the latency and cost budget the product needs. That differentiator is heavily correlated with production miles, not with framework familiarity.

Generalist recruitment flows do not work for this role. LinkedIn searches for 'LLM engineer' return a sea of self-titled candidates whose CVs cite `openai.ChatCompletion.create()` as 'production LLM experience'. Our pipeline is built to filter that out in the first 15 minutes, not in the third interview round.

We source across five channels specific to AI engineering: the open-source contributor graphs of llama.cpp, vLLM, LangChain, LlamaIndex, and OpenAI's Evals repo; papers-with-code leaderboards on the MTEB, HELM and Big-Bench splits relevant to our clients' stacks; Hugging Face Spaces authors with sustained traffic; Kaggle NLP competition top-500 in the last 18 months; and a private network of 640+ CEE AI engineers Nikita built during his time at Neurons Lab, a European consulting shop that delivered 80+ AI projects for EU clients.

Every candidate goes through a 12-minute AI technical interview that probes production signals, not pattern-matching: 'walk me through the last production incident your inference stack had — what did p95 look like before and after?', 'how do you measure output quality for an open-ended generation task on which a human rater is too expensive?'. The AI asks adaptive follow-ups; a human recruiter reviews the transcript and scores before a shortlist lands in your inbox. Candidates who passed our LLM filter in 2025-Q4 had a median of 3 years of post-2022 LLM production experience and a 91% interview-pass rate at our clients.

The last layer is specific to this role: we require every LLM engineer shortlisted to have shipped at least one feature that served >10K users for >30 days. We verify that with a combination of public artifacts (GitHub, HF Spaces, a blog post, a conference talk) and a reference call. If the candidate cannot point us at a production artifact, they do not ship.