BEHIND THE CONCEPT
A better reason
to start a conversation.
What if we could keep learning about exceptional people over time, and recognize when their work gives us something meaningful to talk about?
That’s the idea I wanted to explore with Constellation: a shared, growing picture of people, their ideas, and the connections worth following.
This site is a guided demonstration. The recurring research and suggested connections described below are possibilities for a live version.
HOW THE IDEA COULD WORK
From noticing good work to knowing why to connect.
Start with people and the work that makes them interesting.
A trusted recommendation, a useful open-source contribution, or an influential paper could be the starting point. Constellation would bring that work together with what a recruiting team already knows, so the reason someone stands out stays attached to the person.
Keep learning, even when there is no open role.
Recurring research could revisit permitted sources daily, every few days, or weekly. A new publication might reveal a developing interest. An agreed follow-up might give a recruiter a reason to reconnect. Each discovery adds context; a publication alone does not tell us someone wants a new job.
Show why a connection might matter.
The system would connect a person’s work to questions the company is exploring—and show the sources behind that suggestion. Following those ideas could also lead to collaborators or people in adjacent fields who would otherwise be easy to miss.
Help the right people start a conversation.
A directory of senior technical personas could suggest individual contributors and engineering leaders whose experience and interests connect with the work, explaining why each person might be a good champion. A recruiter could review the reasoning, confirm that person’s interest, and choose whether the recruiter or champion should reach out. A draft gives them a starting point; a human decides what to say and whether to send it.
THE EXAMPLE IN THIS STORY
A shared question can be the opening.
Shopify’s catalogue article discusses the tradeoffs of using AI at scale. FlashAttention-4, coauthored by Tri Dao, explores how to make part of that computing work more efficient.
That gives us a possible shared interest to explore with an internal expert. It is our interpretation of the work, not evidence that Shopify uses the technique or that Tri is open to a role.
LOOK INWARD AS WELL AS OUTWARD
An internal universe of expertise.
We could build a professional persona for every senior technical individual contributor and engineering leader at Shopify. Each would bring together their areas of expertise, past contributions, current responsibilities, and questions they want to explore. These would be grounded in their work and conversations with them, and reviewed by the person they describe.
A search could look for shared meaning, even when a researcher and an employee describe related problems differently. It would compare the external person’s work with these personas, suggest potential champions, and show the relevant experience, shared interests, and when each persona was last reviewed. The recruiter could compare suggestions and confirm who wants to take part.
For Tri, that might mean finding someone who has worked on the speed and cost of running AI models and wants to explore new approaches. This is a proposed internal capability; the demo does not access internal records or represent a real employee persona.
Let the understanding grow.
AI could help with the reading, connecting, and first drafts. People would check the sources, correct misunderstandings, and add what they learn through conversations, with permission. Over time, that shared understanding could make the next introduction more relevant and more personal.
A little more on the technology
A first version could combine scheduled research, a shared store of evidence, a directory of senior technical personas, and a review screen for recruiters. Each connection would keep its supporting link, date, and review history so someone could check how the system reached its suggestion.
Research would use permitted APIs, feeds, or other approved access, with limits on frequency and spending. AI could suggest links and draft text; software rules would control access and flag duplicates, and uncertain matches would wait for human review. Corrections would be retained so repeated suggestions do not become accepted facts. Semantic search could find personas with related experience; access rules would limit which records it can search, and each recommendation would point to the work and interests that support the match. People would review and update their personas as their work evolves.