tags: ["opensource", "ai", "fitness-tech", "startup"]
The Bottleneck Is Not AI
Every physical store owner knows their business needs AI. The problem is not the technology.
The problem is the translator between "I know my business" and "AI knows what to do."
In the industry, this translator is called an FDE — Field Deployment Engineer. You pay ¥20,000/month for one. They serve 5-10 stores max. They quit within 18 months because the job is travel-intensive, repetitive, and boring.
There are hundreds of thousands of physical stores in China. Every one needs an FDE. Almost none can afford one.
MAFBE is that FDE, re-built as open-source AI.
What MAFBE Actually Does
MAFBE — Multi-Agent Fitness Business Entity — is an open-source framework that deploys 9 autonomous AI agents into a physical gym. Once deployed, the agents handle:
- Member management — Tracking, retention prediction, churn alerts
- Class scheduling and dynamic pricing — Capacity-optimized
- Employee evaluation — Objective metrics, not gut feeling
- Supplier management — Automated RFQ and comparison
- Content production — Daily social media output without human writers
- Infrastructure monitoring — Self-healing server and IoT health
- Cost optimization — Tracking every yuan against budget
- Investor narrative — Real-time positioning and reporting
- Risk audit — Independent verification layer checking all agents
The founder does not intervene. The agents communicate through an event bus, cross-validate through an audit agent, and escalate to the founder only when the autonomous system reaches a dead end.
107 Days in a Real Gym
This is not a simulation. MAFBE has been running ZWISERFIT, a physical gym in Dongguan (Wanjiang), China — a neighborhood with one of the lowest fitness penetration rates in the country — since April 2026.
Key production metrics:
| Metric | Data |
|---|---|
| Autonomous runtime | 107 consecutive days |
| Zero-downtime streak | 34 days (Jun 10 - Jul 14) |
| Agents deployed | 9 concurrent AI agents |
| Self-healed bugs | 6 out of 7 critical infrastructure failures |
| Structured knowledge | 1,830 entries auto-generated from operations |
| Production constraints | 39 rules auto-encoded from errors |
| Monthly operating cost | ¥3,369 (target: ¥700/month via optimization) |
The choice of Dongguan Wanjiang was deliberate. If the system works in a neighborhood where fitness culture is at its lowest, it will work anywhere.
Every agent action is logged. Every error is traced. Every recovery is structural — when a bug happens, a new constraint is added so the same bug cannot happen twice.
Why "Deployment > Model"
The AI industry obsesses over models. Which model is best? Which benchmark is highest?
MAFBE obsesses over deployment. Because a model without deployment is a demo. A model with 107 days of production deployment is a business.
What a model can do — generate a fitness article, analyze a trend, write a marketing headline.
What deployment can do — run a real gym for 107 days without human intervention, self-heal 6 out of 7 infrastructure bugs, extract 1,830 structured knowledge entries from daily operations.
When a competitor copies our model stack, they get a demo. When they copy our deployment, they get 107 days of production data — and by then, we are already at day 200.
The Layer Architecture: Built for Real Operations
MAFBE is not a monolithic "AI agent." It is five interconnected layers, each with a specific role:
⚡ Execution Layer — 9 agents running operations (OpenClaw + MCP protocol)
🧬 Memory Layer — 1,830 knowledge entries, 39 constraints (ChromaDB + Mem0)
🎯 Orchestration Layer — Task routing, priority scheduling (LangGraph)
🔁 Governance Layer — Full-trace audit, OPA policy engine (Langfuse)
🧠 Brain Layer — Adaptive model routing (DeepSeek V4 + Qwen)
T-shaped flywheel: The five layers communicate vertically (each layer informs the one above and below) and horizontally (agents within each layer cross-validate each other's actions). This is not a pipeline — it is an operating room where every participant responds in real-time to what each other discovers.
Any competitor can copy one layer. They cannot copy the T-shaped cross-validation network that emerged from 107 days of live operations.
Open Source Model: WordPress for Physical Retail
| Component | License | What it includes |
|---|---|---|
| Agent framework | MIT | Agent configs, Momo scheduler, Wiki engine, API interfaces, hardware specs |
| Store tacit knowledge | Closed | 1,830 entries, 39 constraints, 107-day audit logs — unique to each store |
| Premium hosting | ¥39/day | Zero-ops, automatic upgrades, hardware ecosystem access |
Open source in the MIT layer. Closed in the store-specific layer.
Why? The framework is universal — any gym can run it. The tacit knowledge is the store owner's 7-year operational experience encoded as AI decision patterns. That is their digital asset, not ours.
Think WordPress.org vs WordPress.com. The code is free. The hosted experience is the product.
For Investors: The Numbers That Matter
Now (Seed stage): MAFBE = AI FDE at ¥39/day, replacing ¥20K/month human FDE. ¥39/day × 10,000 stores = ¥140M ARR. No hardware, no data moats needed.
12-18 months (Series A): MAFBE + KinTwin hardware. Hardware cash flow + behavioral data layer.
24-36 months (Series B+): MAFBE + KinTwin + Zeus Protocol. Cross-store data interoperability protocol. Protocol fee revenue.
Investors do not need to believe in the full vision. They just need to believe that hundreds of thousands of physical stores need an affordable FDE — and MAFBE is the first AI version of one.
Get Started
git clone https://github.com/ZWISERFIT/mafbe-fde.git
cd mafbe-fde
./deploy.sh --store-id "YOUR-GYM"
10 minutes. 9 AI agents. One running gym.
ZWISERFIT is an AI-human co-symbiotic platform. AI handles standardized operations; humans handle temperature and value. All behavior data belongs to the user — DID-secured, MPC-protected, chain-verified.
Top comments (0)