Get in Touch
Hire Talent
Tell us the role
Hire AI & ML Engineers

AI Engineers Who Ship to Production, Not Notebooks

A shortlist in 72 hours, contributing inside two weeks, and a 30-day guarantee if the fit is wrong.

72-hour shortlist · 30-day replacement guarantee · Month-to-month

Three Different Jobs, One Job Title

The most expensive hiring mistake in AI is not knowing which of these you need.

The ML researcher builds and trains novel models. Deep mathematical background, publishes, works in notebooks. Genuinely valuable if you’re pushing the state of the art. Most companies aren’t.

The ML engineer takes models — often someone else’s — and makes them run reliably at scale. Deployment, monitoring, retraining pipelines, latency and cost management. Less glamorous, far more commonly what’s actually needed.

The LLM application developer builds products on top of existing models. RAG architectures, prompt design, evaluation, vector search, agent workflows. This is the fastest-growing category and the one most current requests actually describe. See hire LLM engineers.

Hire a researcher for an LLM integration and you get elegant experiments and no shipped product. Hire an application developer for a genuine research problem and you get a wrapper around an API that doesn’t solve it.

We spend most of the discovery call on this question. Getting it right matters more than anything else on this page.

Screening

What We Screen For

Production deployment, not notebooks

Has their work run somewhere real, under load, with monitoring? The single most useful filter in this category.

RAG architecture

Chunking strategy, embedding choice, retrieval evaluation, and knowing why a naive implementation returns confident nonsense.

Vector databases

Pinecone, Weaviate, pgvector. And the judgement to know when you don't need one.

Evaluation discipline

How do they know the system got better? Teams without an evaluation harness are guessing, and it shows within a month.

Cost management

The real differentiator. LLM bills surprise people badly. Caching, model routing, token discipline, and knowing when a smaller model is sufficient.

MLOps

Versioning, reproducibility, monitoring for drift. Matters as soon as anything is live.

What This Costs You When It Goes Wrong

The characteristic AI failure is not a system that breaks. It’s a system that works convincingly and is wrong often enough to matter.

A support assistant retrieves the wrong section of a policy document and answers with complete confidence. There is no error, no exception, no alert. The answer is fluent and incorrect. If it’s customer-facing, you find out through complaints. If it’s internal, staff quietly stop trusting it and go back to asking a colleague — and you’ve paid for a system nobody uses.

The second cost is financial and arrives faster than people expect. A retrieval step that stuffs 30,000 tokens of context into every request, called on each page load, with no caching. It’s fine in testing with five users. At a thousand users a day it’s a bill that gets escalated to your CFO. The fix — caching, tighter retrieval, routing simple requests to a smaller model — is straightforward, but only if someone thought about it before launch.

Both failures come from the same gap: no evaluation harness. Without a test set and a score, every change is a guess, and “it seems better” is the only available measurement.

Seniority, Defined

Mid-level (2–4 years). Implements against a defined architecture. Comfortable with the major APIs and frameworks. Needs direction on system design and evaluation.

Senior (4–7 years). Designs the system. Makes the retrieval, model, and infrastructure decisions, builds the evaluation harness, and manages cost deliberately rather than reactively.

Lead (7+ years). Owns AI direction, makes build-versus-buy calls, handles the honest conversation about what AI will and won’t solve for your business.

Note that seniority bands run shorter here than in other stacks — the field is young, and five years of relevant experience is genuinely senior.

Common Requests

RAG over internal documents. Search and question-answering across company knowledge. The most frequent request by a wide margin, and the one most often underestimated.

LLM features in an existing product. Summarisation, drafting, classification, extraction. Usually straightforward to prototype and harder to make reliable.

Model fine-tuning. When prompting isn’t enough. Less often necessary than people assume — worth having someone tell you honestly.

Recommendation and ranking systems. Classical ML, still very much in demand, often more valuable than an LLM feature.

Prototype to production. A demo that works and needs to become a system that holds up. Extremely common right now. Related: AI integration and automation.

What a Good First Month Looks Like

Week one. They ask about your data before your model. Where does it live, who owns it, how often does it change, is it clean enough to retrieve against. Expect a good hire to spend the first week establishing a baseline and building a small evaluation set — thirty to fifty real questions with known-good answers. That set is unglamorous and it’s the foundation for everything after.

Weeks two to four. They ship something narrow and measurable rather than broad and impressive. One use case, evaluated, with retrieval quality reported as a number. Expect them to tell you which questions the system handles badly — a candidate who reports only successes in month one is either not measuring or not telling you.

The warning sign is a demo with no numbers attached. An impressive notebook that answers three cherry-picked questions is not progress. Ask what the score was, on what set, and how it changed this week. If those questions don’t have answers by week four, the project has no way to improve.

Mistakes We See

Starting with fine-tuning. It is the first instinct and it is almost always wrong. Better retrieval and better prompts solve most problems more cheaply, and a fine-tuned model locks you to a base model you’ll want to change within a year. Fine-tune when you need consistent structured output or genuinely specialised domain language — not to teach the model facts, which retrieval does better.

Treating RAG as a solved pipeline. Chunk, embed, retrieve top-k, generate. Naive implementations fail on documents with tables, on questions spanning two sections, and on anything needing a date filter. Chunking strategy and reranking matter more than model choice, and that’s where an experienced engineer spends their time.

Hiring for AI when the blocker is data engineering. If the documents are scattered across three systems, inconsistently formatted, and updated by hand, no model fixes that. We’d rather say so during discovery than place someone into a project that cannot succeed. See hire data engineers.

Starting from the technology rather than the decision. A model is only useful if something changes because of its output — a queue gets prioritised, a document gets routed, a customer gets contacted sooner. Projects that begin with “we should use AI” and look for an application afterwards tend to produce something accurate and unused. Ask which decision improves, and by how much, before anyone opens a notebook.

How It Works

  1. Discovery, 30 minutes. Starting with which of the three roles you actually need.
  2. Shortlist within 72 hours. Three to five profiles matched to that specific shape.
  3. You interview. Your process, your technical test.
  4. Onboarded in two weeks. Contracts, NDA, IP assignment, and access to data and training environments.
  5. 30-day guarantee. One email, replacement candidates within 48 business hours, no cost. Full terms →

Rates depend on seniority, problem domain, and engagement length. Tell us the role and we’ll give you a firm number.

Tell us the role and we will send a shortlist within 72 hours.

Tell Us the Role
FAQs

Questions

Do we need a researcher or an engineer?

If you're building a product on top of existing models, you need an engineer or an application developer. Research roles make sense when your competitive advantage depends on a model nobody else has. That's rarer than it sounds.

Can they work with our existing data?

That's usually the first question worth answering, and it's often a data engineering problem before it's an AI problem. If your data isn't accessible and clean, no model will save you. We'll say so during discovery rather than after.

How do we control API costs?

Ask candidates directly during interview. A senior engineer will talk about caching, routing simple requests to smaller models, prompt compression, and setting hard budget limits. Someone who hasn't managed a production bill won't have an answer.

Is our use case actually suited to AI?

Sometimes not, and we'd rather tell you that during discovery than place someone on a project that can't succeed. Rules and search solve a surprising number of problems that get framed as AI problems.

How do we know whether it is working once it ships?

Decide the measure before you build. For most applied work that means a small labelled evaluation set and an agreed threshold, checked on a schedule. Teams that skip this cannot tell an improvement from a regression, and end up arguing from anecdotes about individual outputs.

Get Your Shortlist

Tell us the role. Three to five profiles within 72 hours.