PriceCharm

PriceCharm lets a service business train an AI to quote jobs the way the owner would. This repository holds the whole product.

It's live at pricecharm.com. Try it as a customer, no sign-up needed.

Meta AI Developer Hackathon (Startups track)

PriceCharm runs on Meta's models through the Meta Model API, in Standard mode only. The API refuses Contributor-mode models, because they would let Meta train on customers' photos and owners' prices.

• Muse Spark 1.3 runs the customer's quote chat, reads old quotes and invoices into past jobs, and drafts each quote from the owner's own prices. See apps/api/src/ai/index.ts.
• Walk-around video quotes. A customer films the job instead of taking photos. Muse Spark 1.2 watches it and hears what they say, then picks what the price depends on, like windows to count or a driveway to size. SAM 3.1 follows each window through the video and counts it once, then outlines what it found on a still for the owner. Each window SAM outlined on that still is cut out onto one numbered sheet, and Muse Spark grades every one up close: what kind it is and what state it's in, like "4 large picture windows, heavy hard-water stains". It also notes the materials and condition that change the work, like heavy moss or high-end pavers. The count becomes the size on the quote, in the unit the owner prices by, and the rest goes to the price as details. See walkaround.ts, meta-media.ts and live/videos.ts.
• Photos are read the same way. Muse Spark looks at all of a customer's photos together, SAM 3.1 finds each thing to count in every photo, and the grading sheet throws out SAM's mistakes and grades the rest. Two photos can show the same window, so a count is at least what one photo shows and at most all of them added up, with Muse Spark's own count deciding between. Like a video, it's read outside the request: the chat shows the customer it's looking, and the Charm carries on from what the photos show once they're read. See live/photo-readings.ts.
• Muse Image edits the customer's own photo to show the job done, like the same patio with the moss gone. The quote shows the two as a before-and-after the customer can drag across. Owners can turn it off. See after-preview.ts and live/previews.ts. Muse Image also makes the photos that come with the practice customers owners train on (practice-images.ts).
• WhatsApp. The same quote chat runs on a business's own WhatsApp number, photos and videos included. Owners connect their number in Settings through Meta's Embedded Signup, and a quote they approve after WhatsApp's 24-hour window goes out on a template. Other businesses can connect once Meta's App Review is through. See channels/whatsapp.ts and channels/conversation.ts.
• Muse Voice Transcribe turns a customer's WhatsApp voice notes into text, so they can talk through the job instead of typing it. See voice.ts.
• 21 languages. A customer picks a language or just writes in one, and the chat and quote follow. Fixed words are translated ahead of time, and Muse Spark translates the rest, like quote lines and what a video showed. The job itself stays in English, so the price doesn't depend on the language. See languages.ts and translate.ts.
• Customers' own agents. An AI like Muse can ask a business's Charm for a quote on someone's behalf and send photos, and it gets a price within the owner's rules. It can't say yes for them: it gets a link the person opens. See live/agent.ts, and agent-demo.ts for Muse on the customer's side.

What we built during the hackathon

• the setup flow: pick a trade, add past jobs (with invoice reading), and practice on made-up customers until the AI prices like the owner
• the pricing engine, which learns rates, minimums and price style from the owner's own jobs, and checks every draft against them
• the customer quote chat, on a link and as a website widget, where every customer sees a price and anything unusual goes out as a range the owner can check
• the owner app: quotes waiting for an OK, replies, settings, and what the AI knows
• plans and billing rules, the public site and live demo, and the production deploy
• Tell your Charm, where an owner changes prices for a busy stretch in their own words
• QuickBooks invoice import, waiting on Intuit's review for production
• everything on Meta above: the Muse provider, walk-around video and photo reading with SAM 3.1, Muse Image's after pictures and practice photos, WhatsApp, voice notes, the 21 languages and the agent API

Folder	What it is
apps/api	Express API, MySQL/MariaDB storage, AI integrations, billing, email and the background worker
apps/web	The Next.js web app: the public site, setup, the owner's app and the customer chat
packages/trades	The trades and their services, shared by the API and the web app
docs	How pricing works and operations guides (database, AI keys)
scripts	Local development, test and setup scripts

Get started

You need Node 24 (.nvmrc) and Docker, for PriceCharm's own local database.

npm ci
npm run db:start
python3 scripts/setup-local.py
npm run migrate --workspace @pricecharm/api
npm run dev

Open http://localhost:3100. The API runs on port 4100.

Add PriceCharm's own development AI keys to apps/api/.env (AI keys guide), or set AI_PROVIDER=demo for deterministic, labelled demo responses.

Everyday commands

All commands run from the repository root.

Command	What it does
npm run dev	API, worker and web app. All three reload on save; restart for .env changes.
npm run typecheck	Typechecks every app
npm run build	Builds every app
npm test	API tests that don't need a database
npm run test:db	Every API test, including database tests, against a throwaway quoting_test database
bash scripts/e2e.sh	Playwright suite against a real API and the test database, on ports 3300 and 4200

Safety checks

The API refuses to start a staging or production server with settings that are only safe on a laptop:

• demo AI, local email or local payments
• missing APP_ENV, TRUST_PROXY or provider settings
• an unencrypted connection to a remote database
• email codes in API responses
• a test Stripe key in production, or a live one in staging
• no Turnstile secret, or Cloudflare's test secret in production
• no SENTRY_DSN in production

Production runs on one EC2 instance, with MariaDB 11.4 in Docker beside the app. It has a rows-only app user and a separate migration user (database guide).

CI and deployment

• CI (.github/workflows/ci.yml) runs on every pull request and every push to main. It covers the API's formatting, types, database tests and build, and the web app's formatting, types, build and end-to-end suite.
• Deploy API and Deploy web run on pushes to main that change their app. When the repository variable ENABLE_DEPLOY is true, they copy the release to the server over SSH and restart it with PM2.

The full runbook is apps/api/deploy/LAUNCH.md.