AM2TECH LLC

Case study · SaaS platform

CurlDash

A multi-tenant platform that runs curling clubs end to end: members and families, leagues and playoffs, events and bonspiels, waivers, payments, a store, a public website for every club, and native mobile apps. Designed, built, and operated by AM2TECH, from the first sketch to production.

4
clients on one API: web app, public club sites, admin console, mobile
500
REST endpoints across 85 controllers
130+
zero-downtime database migrations
~4 mo
from first commit to live in production

The problem

Curling clubs are run by volunteers juggling spreadsheets, email lists, paper scorecards, and a separate payment tool for every season. Draw schedules change, subs are found by group text, and scores get typed in hours later, if at all.

What we built

One sign-in that covers every club a curler belongs to. Leagues with schedules, standings, tiebreakers, and playoff brackets. Event registration with team entries and check-in. Sub requests, family accounts, signed waivers, dues, and card payments that go straight to each club's own Stripe account, with exports for the club's books. Livestream links on every game, and public bonspiel pages where a visiting team finds its games without signing in. And a phone camera that reads the scoreboard at the end of a game.

Tech stack

Backend

  • Java 25
  • Spring Boot 3.5
  • Spring Security
  • Spring Data JPA
  • MySQL 8
  • Liquibase
  • Maven

A single stateless REST API behind every client, with JWT auth, role-based permissions, and tenant scoping on every request.

Web

  • Angular 22
  • Signals
  • Zoneless
  • Angular SSR
  • Web Awesome
  • TypeScript

Standalone, OnPush components on the latest Angular. Club websites are server-rendered for speed and search, each on its own subdomain with its own theme.

Mobile

  • Ionic 8
  • Capacitor 8
  • Android
  • iOS
  • FCM push
  • Biometrics

Native apps sharing the same API and sign-in, with push notifications, deep links, Face/fingerprint unlock, and an offline mode for game day.

AI

  • Claude (Anthropic)
  • Vision
  • Anthropic Java SDK
  • Structured output
  • MCP

Reads a photo of a physical scoreboard and turns it into end-by-end scores (details below), and a private MCP server lets the operators work with live data through Claude.

Integrations

  • Stripe Connect
  • Resend
  • Twilio Verify
  • Firebase
  • iCalendar
  • Inbound email

Payments, transactional and inbound email, SMS sign-in codes, push notifications, and calendar feeds that sync to any phone.

Infrastructure

  • Docker Compose
  • Caddy
  • GitHub Actions
  • GHCR
  • DigitalOcean
  • Cloudflare

Every push to main builds versioned images and deploys them automatically. Automatic HTTPS for every club subdomain.

Applied AI

Reading a curling scoreboard from a photo

A curling scoreboard is not a table of numbers. Teams hang numbered tiles along a fixed score scale: the number on a tile is the end just played, and the column it hangs in is the team's running total. Reading it takes a minute even for people who know the game, so scores often never made it online.

In CurlDash, the skip takes a photo with their phone. A few seconds later the app shows the full end-by-end line score, both totals, and who had the hammer, ready to check and save.

  • About 6¢ a photo. Under 50¢ for a full eight-game league night, measured in production.
  • Cost guardrails. Limits per game and per club per day stop a read before it's sent, so a stuck button can't run up a bill.
  • Human in the loop. The result is a preview. Nothing is saved until the skip confirms it, and they can swap teams or fix an end.
  • No competing product does this. We found no shipping curling software that reads physical scoreboards.

How the prompt was engineered

The first version asked the model to read the board and do the arithmetic. It was close, but "close" is wrong in a scoring app. The working design splits the job: the model does perception, and code does the math.

  1. Prepare the image

    The client downscales the photo on the device to the vision model's useful resolution, honoring camera orientation, so uploads stay small and fast.

  2. Describe the structure

    The prompt teaches the board's layout: three rows, a hammer column to skip, a score scale from 1 to 14, and what a tile's number and position each mean.

  3. Ask only for what it sees

    Instead of scores, the model returns a strict JSON grid: for each score column, which end tile hangs above and below. A worked example in the prompt anchors the format.

  4. Let it think

    Claude runs with adaptive extended thinking and a token budget sized so reasoning never crowds out the answer.

  5. Validate in code

    Java turns the grid into end scores (column minus previous column), drops physically impossible tiles (a negative end or more than 8 points), derives the hammer, and trims unplayed ends.

  6. Fail safely

    Prose instead of JSON, an empty grid, or a refusal all become a clear "No scoreboard detected" message, and an empty result can never wipe scores already entered.

The pattern carries over to any document or photo workflow: let the model extract what it observes into a strict schema, keep the business rules in deterministic code, check the output against what is physically possible, and keep a person in the loop for the final save.

Engineering highlights

Multi-tenant by host name

One deployment serves every club. The host decides the surface: app, admin, landing page, or a club's own server-rendered site at its own subdomain, with data isolated per club on every query.

Passwordless auth and "act as"

Sign-in by emailed or texted code, rate limited. Parents manage their kids' profiles through an acting token that keeps the real signer as the subject for audit, and site admins can impersonate safely for support.

Fine-grained roles

Members can hold several roles whose permissions combine, plus scoped powers: league and event managers, and skips who can score their own games and manage their team's subs.

Marketplace payments

Stripe Connect direct charges land in each club's own account, with a platform fee that respects a yearly cap, held spots that release after 30 minutes, refunds, receipts, dues, add-ons, per-club test mode, and accounting exports with bank payout matching.

Offline game day

Curling rinks often have poor signal. The mobile app caches schedules and games, queues scores entered offline, and checks for conflicts before syncing.

Leagues, playoffs, brackets

Schedule builders, standings with tiebreakers, reusable playoff templates, multi-draw brackets, winner-only results where clubs don't keep end scores, and a "My Path" view that projects a team's route to the final.

Notifications that respect people

Push and email per topic, per person, sent only after the database commit so nobody gets a score that was rolled back. Game reminders go out at 8 AM in the club's own time zone.

Safe content and documents

News posts can be emailed in, even from Mailchimp. All HTML is sanitized on save and on every read. Signed waivers are rendered to PDF and kept on a retention schedule.

Game day for spectators

Livestream links set per sheet camera, league, draw, or game (the most specific wins), a live board, and public bonspiel pages where a team finds its lineup, games, and playoff path, no account needed.

Claude on production, safely

A private MCP server lets a few named admins ask Claude about live data and apply fixes: OAuth 2.1 sign-in with PKCE, read-only queries, and every change previewed, row-capped, approved by a person, and audited.

Privacy and account lifecycle

Self-service account deletion with a grace period, anonymized history, audit logs for admin actions, and an email log for support.

How it was built

CurlDash was built by a small senior team using AI-assisted development throughout: written conventions the tools follow, small reviewed changes, and every change compiled and tested before it ships. That is how one team delivered a product of this size in months instead of years.

Every database change is a new forward-only migration, applied automatically on deploy, so releases never need hand-run database changes.

What it shows

  • Product design from a blank page to a live service
  • Full-stack Java and Angular on current versions
  • Native mobile apps with offline support
  • Practical, cost-measured AI in a real workflow
  • Payments, compliance, and privacy done properly
  • Automated build, deploy, and operations

Have a project like this?

Whether it's a platform, a mobile app, or a place where AI could save your people real time, we'd like to hear about it.

contact@am2tech.llc