USMMA Wrestling: AI-powered site management for non-technical stakeholders
Built a recruitment platform for the US Merchant Marine Academy wrestling program — stakeholders describe changes in plain English, a Claude Code sandbox makes the edits, creates a PR, and returns a Vercel preview link for approval.
Posted by
Gene KobilanskyRelated reading
GMB Praxis: adaptive fitness app for physical autonomy
GMB is a tech-enabled fitness company that helps its clients achieve the foundations of physical autonomy: strength, flexibility, and control. We built and optimized GMB Praxis, a web app with custom adaptive sessions, flexible scheduling and durations, and seamless video streaming.
Samsung Galaxy: interactive Amazon store experience
Samsung wanted a sleek, interactive Amazon Store experience to highlight their Galaxy product lineup, integrating 360° views, immersive storytelling, and platform consistency across web and Amazon Fire.
Interactive Amazon Store Experience for Temptation's Cat Treats
Temptation's Cat Treats is a popular pet food brand that specializes in cat treats. Their mission is to engage cat owners and offer tasty, fun treats for their feline companions. For this project, the goal was to create an engaging, playful mini ad site within their Amazon Store to boost user interaction and product awareness.
Objectives
- 🙋🏻♀️ Build a recruitment site for the USMMA wrestling program
- 🙋🏻♀️ Let stakeholders update the site without contacting a developer
- 🙋🏻♀️ Keep changes safe — preview before going live
- 🙋🏻♀️ Minimize friction with passwordless authentication
Results
- ✅ Live recruitment site serving the USMMA wrestling program
- ✅ Stakeholders update the site by chatting — no code, no tickets
- ✅ Every change creates a PR with a Vercel preview link for approval
- ✅ Magic-link auth — no passwords to remember or reset
Highlights
Chat-to-PR Workflow
Stakeholders describe what they want changed in plain English. The system spins up a temporary Claude Code sandbox, makes the edits, creates a GitHub PR, and returns a Vercel preview link — all within the chat. They review the preview, and if it looks good, the changes go live.
Sandboxed Code Execution
Each edit request spins up an ephemeral Vercel Sandbox — a Firecracker microVM running Claude Code. The AI reads the current site, makes the requested changes, commits to a branch, and pushes. No changes touch production until the stakeholder approves the preview. Safe enough for non-technical users to trigger updates daily.
Magic-Link Authentication
No passwords, no friction. Users enter their email and receive a secure sign-in link (JWT with 15-minute TTL). Sessions persist for 7 days via Upstash Redis. The entire auth flow takes seconds — critical when you want stakeholders to actually use the tool instead of emailing you.
Challenges and Solutions
Mobile Session Reliability
Mobile browsers aggressively kill background connections. I built reconnection logic that detects stale sessions, re-establishes SSE streams, and recovers chat history — so conversations survive phone sleep without losing context.
Making AI Edits Trustworthy
The key insight: never let AI push to production. Every change goes through a PR → preview → approve flow. Stakeholders see exactly what changed before anything goes live. Rate limiting (10 requests/hour) and session management prevent runaway costs.
Beyond the day job
As an assistant wrestling coach at USMMA, I was the bottleneck for every site update. Now the coaching staff can update photos, copy, and layout themselves — just by describing what they want. This is the kind of tool I wish existed everywhere: AI that handles the implementation so domain experts can focus on their actual job.