How do you source AI and machine learning engineers in 2026?

According to Second Talent's 2026 AI Talent Shortage report, there are currently 1.6 million open AI positions globally with only 518,000 qualified candidates, a demand-to-supply ratio of 3.2:1. AI job postings grew 78% year-over-year in 2025 while the qualified talent pool expanded by just 24%. ManpowerGroup's 2026 Talent Shortage Survey found 72% of employers now report difficulty filling these roles.

That gap doesn't close through traditional sourcing. Most ML engineers aren't on LinkedIn waiting to hear from a recruiter. The ones who are actively job-hunting are already fielding multiple offers. The engineers who can actually do the work — the ones building production ML systems, fine-tuning models, writing distributed training pipelines — are mostly invisible to standard Boolean search.

Here's where to find them, what signals to look for, and what actually works when you reach out.

Where ML engineers spend their time online

LinkedIn is where ML engineers list job titles. They build credibility elsewhere.

GitHub is the closest thing to a portfolio this field has. Look for contributions to foundational ML libraries: PyTorch, TensorFlow, JAX, Hugging Face Transformers, LangChain, scikit-learn. An engineer who has merged a PR into any of these repositories has been vetted by the project's core maintainers, which is a stronger signal than any resume keyword. According to Indeed, the machine learning engineer title has seen 53% growth since 2020, but most of that growth is self-reported. Actual merge history isn't.

Kaggle is built around ML competitions with fully public leaderboards. Grandmaster and Master tier competitors are verifiable by performance on real problems, not credentials. Many don't maintain a complete LinkedIn profile. Some don't have one at all.

Hugging Face now hosts over 780,000 open-source ML models. Engineers who have published models there (particularly those with meaningful download counts) are demonstrating measurable output in public, which is the kind of signal that holds up in technical screening.

arXiv (specifically the cs.LG and cs.AI categories) is where researchers publish pre-prints. If someone has authored or co-authored a paper there, they are deep in the field. ATS systems don't index it. Most sourcing tools don't either.

What contribution signals tell you about ML engineers specifically

Not all GitHub activity in ML is worth your time. Someone can star 400 repositories and watch videos about PyTorch without writing a single line of production code.

The signals that hold up: contributions to actively maintained ML libraries, where code passes review by engineers who know the domain. The pull request conversation is often more informative than the code itself. Does the engineer understand tradeoffs? Do they ask architectural questions? Do they engage with feedback constructively?

Notebooks on Kaggle or GitHub show how someone approaches a full problem: data exploration, feature engineering decisions, model selection reasoning, evaluation design. A strong notebook is essentially a technical interview artifact hiding in public view.

Model cards on Hugging Face (the documentation engineers write about their own models) reveal something about communication ability. ML engineers who can't explain their models to a non-specialist tend to be difficult to integrate into cross-functional teams.

Second Talent's 2026 data shows LLM-specific expertise demand has risen 340% since 2023. That concentration matters for sourcing: the engineers with the most relevant skills are spending time on LangChain, LlamaIndex, and related repositories, not on general-purpose Python projects. What commit history actually tells you about a developer covers how to read these signals across a full GitHub profile.

Outreach that works with this cohort

InMail response rates for ML engineers run 3–8% on baseline sends, according to Expandi.io's 2025 State of LinkedIn Outreach report. With personalization tied to the candidate's actual work, rates climb to 10–15%. A 2024 TalentBoard survey found 86% of candidates ignore generic recruiter messages.

The pattern that consistently outperforms: reference the specific thing they built. "I saw your PR to PyTorch's distributed training module" is a different opening than "Your background caught my attention." One requires you to have actually looked at their work. The other signals that you haven't. ML engineers are technical people who can tell the difference in the first sentence.

One other thing worth knowing: Stack Overflow's 2024 Developer Survey found that AI and ML specialists consistently rank meaningful problem-solving above compensation when evaluating opportunities. Leading with the problem your company is working on, before getting to compensation or tech stack details, tends to outperform compensation-first messaging for this cohort. How to write a cold recruiting email a developer will actually respond to has a framework that maps directly to this approach.

How contribution data changes what's searchable

Standard sourcing tools index profiles and resumes. riem.ai indexes GitHub event data (actual contributions, PR activity, repository involvement), which gives recruiters a way to search by what engineers have built, not what they've listed.

For ML recruiting specifically, the most telling signals don't appear in resumes or LinkedIn profiles. They live in version control. A search for engineers with recent contributions to PyTorch or Hugging Face Transformers surfaces candidates who are current and active in the field, not engineers who were active several roles ago. How to source passive software engineers who aren't on LinkedIn covers the broader framework these searches fit into.

Frequently asked questions

How long does it typically take to hire a machine learning engineer?

For specialized roles (NLP, computer vision, reinforcement learning), time-to-fill runs 90–120 days at companies without an established ML research reputation. Sourcing alone takes 3–5 weeks for active pipeline development. Technical screening adds another 2–3 weeks, and interview loops for senior ML candidates rarely run fewer than 4–5 rounds. Starting 3–4 months before the target start date is a realistic planning assumption.

Where are ML engineers geographically concentrated in 2026?

The Bay Area, Seattle, and New York remain the largest pools. Significant secondary concentrations exist in Toronto and Montreal (world-class AI research programs at U of T and Mila), London and Berlin (strong university pipelines), and Warsaw and Kraków (Eastern European mathematical training traditions with growing ML community presence). Remote-first companies have access to the widest candidate pool.

What's the difference between an ML engineer and a data scientist when sourcing?

ML engineers build and deploy production systems: training pipelines, inference infrastructure, model serving at scale. Data scientists typically focus on analysis, experimentation, and model development without the production engineering component. The distinction matters for sourcing because their community platforms differ (data scientists skew toward Kaggle and notebooks; ML engineers toward GitHub and library contributions), their portfolios look different, and what they value in a role differs significantly.

Is Kaggle ranking a reliable signal for production ML work?

It's a strong signal for modeling skill and problem-solving under constraints, but an incomplete one for production roles. Kaggle competitions don't require deployment, scalability, or collaboration with other engineers on long-running codebases. A Grandmaster competitor who also has GitHub contributions to production ML infrastructure is a much stronger combined signal than either alone.

Should you post ML roles publicly or source directly?

For senior and specialized ML roles, direct sourcing consistently outperforms job postings. Senior ML engineers at the experience level most hiring teams are targeting receive an average of 3–5 inbound recruiter contacts per week. A job posting adds to that noise. Direct outreach referencing specific work cuts through it. Save public postings for building inbound pipeline, but don't expect them to close senior roles.