Rxd

Inspiration

Gyms compete on equipment, interiors, and price — an endless capital arms race nobody wins. But members don't quit because the dumbbells got old. They quit when their sense of belonging quietly fades, often weeks before they actually cancel. And operators almost never see it coming until the payment stops.

We asked a simple question: what if an AI agent could watch a fitness community's heartbeat and tell the operator exactly who to reach out to this week — and why — before they're gone?

What it does

Rxd is an agentic retention layer for group-fitness gyms (starting with CrossFit). It turns everyday workouts into community signals, then acts on them through a loop of AI agents:

• Whiteboard agent (vision + OCR): A coach snaps one photo of the gym whiteboard. The agent reads the WOD, structures it, logs each member's result, and infers daily attendance — with zero new behavior from members or coaches.
• Nutrition agent: Members photograph a meal; the agent recognizes the food and estimates macros, removing the friction of manual diet logging.
• Retention agent (the core): It continuously watches engagement signals — attendance, workout records, cheers, event participation — detects members going quiet before they churn, and drafts personalized re-engagement messages for the operator to approve and send.
• Team-season agent: Scores team competitions, tallies individual contribution, and distributes points/rewards automatically.

The result: members keep showing up, and the operator gets an autonomous "who to care for this week" action list instead of manual spreadsheets and guesswork.

How we built it

• Agent orchestration: [function-calling LLM with tool use — e.g., OpenAI / Qwen]
• Vision & OCR: [whiteboard parsing + meal recognition]
• Backend / data: [Node/Python, Postgres, vector store for member context]
• Frontend: [operator dashboard + member app]
• Sponsor tech: [list the Build Week sponsor tools you actually used]

Challenges we ran into

• Whiteboard photos are messy — handwriting, glare, gym abbreviations — so making OCR reliable enough to trust as an attendance signal was hard.
• Designing the retention agent to be proactive but not spammy.
• Turning soft "community" activity into a hard, operator-trustworthy next action.

Accomplishments that we're proud of

• A single coach habit (photographing the whiteboard) now generates records, attendance, and a social feed at once.
• The agent's value isn't prediction — it's converting signals into a specific action an operator will take.
• Early pilot: members re-activating month over month in a real Seoul training center.

What we learned

• In community products, retention is decided between visits.
• "Agentic" only matters if it closes the loop: detect → draft → act → measure the return.

What's next for Rxd

• Expand from CrossFit to HYROX, hybrid training, group PT, combat sports, pilates, and yoga.
• Close the messaging loop end-to-end and track re-engagement outcomes.
• Sit on top of any booking/payment system — no rip-and-replace.

Built With

• agentic-ai
• computer-vision
• function-calling
• llm
• nutrition-ai
• ocr