Every AI founder I talk to has the same story. They opened a role for an ML engineer, posted it, ran a few LinkedIn searches, and got hundreds of profiles that look right on paper. Three months later the role is still open.
The problem isn’t effort. It’s that the methods we use to find software engineers don’t work for AI talent.
Keywords describe claims, not work
A LinkedIn search matches what people write about themselves. “PyTorch”, “LLM”, “Machine Learning” — today these words are on almost every engineering profile. They tell you what someone has touched, not what they have shipped.
The people you actually need — the engineer who made inference twice as cheap, the researcher whose paper your team keeps citing — often have thin profiles. They are busy building, not updating LinkedIn.
The best people aren’t looking
Strong ML engineers get messages from recruiters every week. Most of those messages are generic, so most get ignored. If your outreach looks like everyone else’s, it disappears in the same pile.
This is why “post and pray” and volume InMail campaigns produce so few hires for AI roles. You are fishing in the pool of people who are visible and available — which is rarely the same as the people who are excellent.
Where AI talent actually shows up
When I source for ML and AI roles, I start from evidence of work rather than self-description:
- Papers — authors on arXiv and at NeurIPS, ICML or ACL in the exact area you need.
- Code — maintainers and serious contributors to the open-source projects your stack depends on.
- Talks — engineers who presented how they solved the problem you are facing.
- Teams — the groups that shipped something comparable, and who was on them at the time.
Each of these is a signal that someone has done the work, not just listed it.
What changes with agents
Mapping all of this by hand used to take weeks per role. Today I use AI agents to do the heavy lifting: turning a job description into concrete signals, scanning papers and repositories, and ranking people by what they have built.
But agents don’t make the decisions. Every profile is reviewed by a human before anyone reaches out, and every message references the candidate’s actual work. Agents handle the volume; judgment and relationships stay human.
The takeaway
If your AI role has been open for months, more of the same sourcing won’t fix it. Change where you look: start from papers, code and shipped work — and talk to people about what they built, not about your job description.