Hire AI/ML engineers who can own the whole AI surface at a product scale-up.

For a Series A or early Series B product scale-up, the AI/ML generalist is the engineer who owns the whole AI surface because no one else on the team has time to. In a given week they might fine-tune a lightweight classification model on Hugging Face, stitch an OpenAI or Anthropic call into the checkout flow behind FastAPI, stand up a pgvector store for semantic search, and write a throwaway eval harness to compare Claude 4.5 against GPT-4.1 for a specific customer-support intent. They are the single point of AI literacy at the company, and they keep the product moving until hiring can catch up.

Demand has moved with the market. The 2025 Stack Overflow Developer Survey flagged that 62% of developers use AI tools in production, but a separate Kaggle community survey from late 2025 showed that only 19% of respondent-shipping companies had an AI-specialist headcount above 3 engineers — the rest rely on generalists. That gap is the generalist's entire job description. Hugging Face's own model-download telemetry tells the same story: the fastest-growing category of downloads in 2025 was not frontier LLMs but small task-specific models in the 100M–7B parameter range, which is exactly what generalists deploy when a 70B frontier model is overkill or too expensive.

We see three types of generalist roles most often: (1) the first AI hire, who shows up at a 20–40 engineer product company and inherits every AI feature that has ever been shipped; (2) the AI-literate product engineer, who joins an existing team of 1–2 AI engineers and widens the surface area they can cover; (3) the bridge engineer at a more mature company, who keeps talking to the classical-ML stack (fraud models, recommendation systems) while the rest of the team specialises into LLM vertical roles. All three need one trait: they are comfortable switching between PyTorch training loops, OpenAI API calls, and a 2am inference-cost spike on the same Tuesday.

A generalist is the right hire when your AI team is 1–3 engineers. Above 3, specialisation starts to pay off — at that point you usually want a dedicated RAG engineer, an evals engineer, and an ML platform engineer rather than one person who half-covers all three. Below 1, you are asking a backend engineer to learn AI on the job, which typically costs you 6–9 months of product velocity. The generalist sits exactly in that middle band, and the middle band is where most of our UK and EU clients currently are.

The generalist role is genuinely harder to screen for than any single specialist role because the bar is horizontal, not vertical. A specialist needs depth in one area; a generalist needs credible shipping experience across four or five. Most inbound CVs to us for this title over-index in exactly one area (a PhD who has only published on attention mechanisms; a backend engineer who has only wired up `openai.chat.completions.create`) and quietly miss the rest. Our first-pass filter is built to catch that imbalance inside the first screen, not the third interview.

We source across five channels specific to AI/ML generalists: Kaggle competition top-1000 finishers in the last 24 months (which preferences people who have shipped end-to-end notebooks, not just read papers); Hugging Face model authors with download counts between 500 and 50,000 (the sweet spot where someone actually trained and published, but is not a full-time researcher); GitHub contributors to scikit-learn, PyTorch Lightning, LangChain, LlamaIndex, and FastAPI repos; attendees of European AI engineering conferences (PyData London, MLConf, AI Engineer Summit) who have a shipping artifact behind their talk; and a private network of 640+ CEE AI engineers Nikita built during his time at Neurons Lab, a consulting shop that delivered 80+ AI projects for European clients across fintech, health, retail, and maritime.

Every candidate goes through a 12-minute AI technical interview that deliberately switches context mid-conversation. A typical session moves from 'walk me through the last time you chose between a fine-tuned BERT and an API call for a classification task' to 'how did you instrument cost-per-request when you first shipped an LLM feature?' to 'show me how you would debug a scikit-learn model whose F1 dropped 8 points in production'. Generalists who have actually shipped handle the switching fine; candidates who have over-indexed on one library (we see a lot of Keras-only or PyTorch-only candidates) stumble on the first hand-off. A human recruiter reviews the transcript and scores before a shortlist lands in your inbox.

The last layer is specific to this role. We require every AI/ML generalist we shortlist to point at three production-shipped artifacts across at least two different AI paradigms — e.g. one classical ML model, one LLM-backed feature, and one lightweight inference or ops piece (a cost dashboard, an eval harness, a batched-inference optimiser). We verify those with a combination of public artifacts (GitHub, Hugging Face, Kaggle, a conference talk, a blog post) and a reference call. Candidates who can only point to one category do not make shortlist.