If you followed our earlier posts, you saw big claims: nine agents, one gym, zero human intervention. Those posts were real attempts, written at the time. This post is about what we learned since — and where we actually are.
The wish. I have run real fitness businesses for ten years; one gym in Dongguan has been open for seven. My dream was simple: let AI be real employees in my own gym — run the store, and turn members' healthy habits into assets that belong to them.
The wall. When we actually tried, today's AI could talk, but it could not be trusted with a real business. Many ideas — MAFBE, our early store-manager Momo, and more — kept waiting on better foundations. The technology was not there yet.
The ground floor. So we stepped back and rebuilt from the bottom: store operations, then behavior capture, and finally the lowest problem of all — agents themselves are not reliable. That is why we built LAO. Without reliable agents, AI will never be able to run a store.
Names came later. We never set out to chase trends. We wanted one thing: turn the model's probabilistic answers into answers that match what users actually know and expect. Only afterwards did we learn the industry has a name for this — post-training. Same with our next two: specialist agents solving specific business problems (that's fine-tuning), and data that belongs to users (that's Web5). Work first, names later.
Where we are today.
- LAO — shipped, open source, our first product: the independent third party between your LLM and your Runtime, on your side.
- Saros — the digital store manager, in preparation. It carries forward Momo, our first store-manager attempt.
- Melody — the metabolism coach, in preparation.
The destination has not changed: AI truly running the store, and data belonging to users. We just learned you have to pour the foundation first.
Official site: zwiserfit.cn (English) · zwiserfit.cn/zh/ (中文)
Investor brief: zwiserfit.cn/for-investors/
Top comments (0)