We run 9 AI agents on 2 CPU cores to operate a real gym. Here's the three-layer architecture that makes it work.
Layer 1: Momo — The Scene Layer
Momo is our open-source store brain. It handles daily operations: face check-in, workout record management, class scheduling, payment processing. Everything a gym needs to run, automated.
Momo has two faces:
- Saros (B2B): A store digital co-manager that replaces the admin overhead of a traditional gym manager. Scheduling, member management, daily reporting — all AI-driven.
- Melody (B2C): A personal metabolic coach that tracks energy, glucose-lipid, and hormone metabolism across three dimensions. Available to every member as their personal AI health companion.
Together they handle 99% of daily gym operations autonomously. The 1% that needs human intuition — member relationship moments, strategic decisions — gets escalated to our human team.
Key design principle: One AI brain, two product faces. Members see "Momo helped me train today." Store owners see "Momo is running the store." They don't need to know it's two different products.
Layer 2: KinTwin — The Technology Core
KinTwin is the verification infrastructure. Edge computer vision running on low-power hardware validates every behavior record at the point of capture — not in the cloud, not after the fact, at the moment it happens.
Why does this matter? Because behavioral data is worthless if it can't be verified. A step counter says you walked 10,000 steps. A survey says you exercised. Neither can prove it.
KinTwin changes that. When a member walks through the door, the system knows. When they finish a workout, the system records it — with biometric binding and hash-chain verification. This is the Proof of Physical Behavior (PoPB) protocol.
The Nourish comparison: Nourish raised $215M proving insurers will pay for health data. But their data comes from user self-reporting — questionnaires people fill out. KinTwin's data comes from hardware — biometric-verified, continuously collected, impossible to fake. Nourish proved the demand. KinTwin delivers the real answer.
Layer 3: Global Operations — The Capital Layer
Zeus Protocol sits at the top, monetizing verified behavioral data through protocol fees — not data sales. This is a critical distinction:
- Data seller model: "Give us your data, we'll sell it to insurers" — you lose ownership, we extract margin.
- Protocol fee model: "You own your data, insurers pay to verify it through our protocol" — you keep ownership, we take a protocol fee for enabling the transaction.
Nourish validated the insurance data path. KinTwin validates the data is real. Zeus Protocol makes it tradeable without compromising user ownership.
Stella, our independent audit agent, runs continuously across all three layers. Every transaction, every behavior record, every asset token — Stella audits them. This is not "trust us." This is "verify us."
The Real Lesson
9 AI agents. 2 CPU cores. 3.6 GB RAM. One gym in Wanjiang, Dongguan — the lowest fitness penetration area in China. Running live since April 2026.
The thesis is simple: AI operates. Humans oversee. Users own their data. And we do it all on hardware that fits in a backpack.
Not "AI helps humans work faster." AI operates the business. Humans provide the judgment. Users own the value they create.
- GitHub (open source): https://github.com/ZWISERFIT/ZWISERFIT
- Dev.to: daily engineering dispatches from a gym run by agents
Top comments (0)