I thought “context engineering” was just a fluff word until I had to do it myself. I think I found a way to make AI think like humans
I spent the past few months trying to solve one problem:
How do we get AI to actually match people in a way that feels human?
How do we get AI to make the right match, like a professional human matchmaker who knows everyone deeply?
How do we get AI to make the right match, like a professional matchmaker who knows everyone and what connects them?
I tested vector databases first. They’re great at “similarity.” But when it came to matchmaking, all I got back was a list of 20 people who were vaguely alike. It was too shallow
I knew that if I wanted to build an AI agent that can autonomously matchmake people properly, I had to find a way to make AI store data like a human
This is where Graph databases came into play. It doesn’t just show similarities. It reveals exactly what is similar and why 2 people would be a good match. for e.g.
- You like hiking → Alice likes hiking.
- Alice did an exchange in Germany → you’re from Germany.
- Alice is in your course → she needs a study buddy.
and once you RAG this graph up (GraphRAG), you’ll have an AI that knows exactly why 2 people out of millions should meet
At scale, vectors become too noisy and very slow meanwhile graphs get sharper with every node and connection
That’s why I’m betting on graphRAG while building Ember, an AI that connects students at UCalgary & UWaterloo and helps them find the right time to meet
Plenty of startups like Boardy, Series, Ditto, and Sitch are exploring this idea across different industries.
Investors are bullish too, with Amber at Patron Fund among the most vocal supporters
I’ll be sharing more insights over the next months as I continue to research this “AI RelationshipOS” space with the help of my professor Christian Jacob from University of Calgary
If you’re exploring this space too, let’s connect. And check out my video (and livestreams) on context engineering where I go deeper into how I’m building Ember (Orbit v2)