An AI embed serving 200k+ active users.
Kleio, a Paris AI startup, sells an AI product its customers embed in their own sites — so the interface has to be fast on a page Kleio does not control, and retrieval has to keep returning the right context as the corpus grows. I joined their engineering team on subcontract as the second engineer and worked on both: the low-latency embed UI, and the Node and Python APIs behind it. Across eighteen months the embed passed 200k active users over their client deployments.
The problem
An embed lives on a page you do not control, next to whatever else that page is already doing. It gets one chance: if it loads slowly, or answers from the wrong context, the visitor closes it and the client notices. Both problems get harder as the corpus and the traffic grow, and they pull against each other — the more you retrieve, the slower you get. I joined as the second engineer while Kleio was already scaling, which meant building for the load the product was about to have rather than the load it had.
What I built
How it works
Results
Running it
The engagement ended in June 2025 and Kleio's team runs it from here — they own the code. The embed is live across more client sites today than when I left. Being the second engineer on a small team is what taught me to treat interface, API, and retrieval as one path rather than three handoffs — which is how I work on my own projects now.
Related
I take on a small number of projects at a time.
Tell me what you're building and where it hurts. I'll say honestly whether I'm the right person for it.