A three-year, two-part relationship. We began with Habuild in October 2023 by building their member-facing product. We are now deploying Sketch as their internal company brain, and Habuild is the first design partner for Sketch. The two builds are the same idea at two scales: a context graph of each member, then a context graph of the whole company.
What we wanted to find out
Habuild runs large guided yoga and meditation programs. A member touches the company in many places: the app, WhatsApp, classes, payments. Could we turn all of that scattered history into one rich, living picture of each member, and put an assistant on WhatsApp that actually understood the person it was talking to?
Build one: Beetu, a graph of every member
The first version of Beetu, Habuild’s member-facing WhatsApp assistant. Underneath it is a context graph of each member: we took every touchpoint and turned it into per-member context, then built the agent that answers over it. That agent now handles over a million queries a month.Roughly 50,000 a day. Built and maintained alongside Habuild's team across about two years, and since transitioned to them.
The assistant is only as good as the per-member context behind it. The effort that paid off was assembling each member’s history cleanly, not prompt-tuning the agent on top.
Build two: Sketch as the organization’s brain
We deployed Sketch, our own context-graph platform, as Habuild’s organization-wide “Org AI”, and Habuild became its first design partner, running it self-hosted on their own infrastructure and helping shape the roadmap. Sketch resolves their internal tools into one company-wide knowledge graph the whole team can simply ask, with automations and human-approval steps layered on top, and each tool’s permissions carried through so people only ever see what they already can.
The hard part was rebuilding the entity graph: deduplicating tens of thousands of entities across six tools, using deterministic and LLM methods together.
What we figured out
People adopt it when it lives where they already chat and they can just ask. The same idea held at both scales: assemble the rich context object once, and the assistant standing on top inherits the benefit.