Pricevane — an AI-powered multi-tenant SaaS for competitor price monitoring

A production-shaped SaaS for small e-commerce teams. It crawls rival storefronts overnight, matches their listings to your products, and tells you the moment one of them goes under you — with the product matched, the movement measured, and the minute it happened.

The hard part is not the crawling, it is knowing that their "ErgoMesh Task Chair, Graphite" is your SKU. That is done by an LLM that returns a confidence and a written reason for every pairing — and never applies one without a human confirming it.

Subscription billing, per-tenant isolation, plan limits and usage metering are built the way a real SaaS has to build them: enforced in the database, not in the interface.

This is a portfolio project. Every company, customer, quotation and figure in the interface is invented, and the shops the crawlers visit are three fictional storefronts built for the purpose. No real retailer is ever crawled. The architecture, the isolation tests and the billing ledger are real.

---

The three decisions worth explaining

Most of this product is ordinary. Three parts are not, and they are the parts that decide whether it survives contact with real customers.

1 · Tenant isolation lives in Postgres, not in application code

Every tenant-scoped table carries an organization_id and a row-level security policy, and the policies are FORCEd so they apply to the table owner as well. There is no path — ORM, raw SQL, crafted filter, forgotten WHERE — that returns one customer's rows to another.

Application-layer filtering was rejected because it fails open: a missing .eq('organization_id', …) in one handler leaks a whole table and nothing complains, because the tests exercise the handler that has the filter. A policy fails closed — get it wrong and you see nothing, which is a bug you find in the first minute rather than in a disclosure email.

tests/rls/ signs in as one organisation, crafts a query for another organisation's rows, and asserts it gets none.

A related trap, found the hard way and now covered by regression tests: a table-level GRANT SELECT silently overrides a column-level REVOKE. Secret columns (notification_channels.secret, organization_settings.api_key_enc) are protected by per-column grants in 0007_column_grants.sql.

2 · Stripe webhooks are idempotent by ledger, not by hope

Every event is claimed in an event ledger inside the same transaction that applies it, keyed on Stripe's own event id. A replay finds the row already processed and stops. Out-of-order delivery is handled separately, with a watermark on the subscription — an event that describes an older state than the one already recorded is discarded rather than applied.

3 · The model never merges anything on its own

Matching returns a confidence and a written reason, and everything it proposes waits for a person to confirm or reject it — at 0.94 and at 0.99 alike. Both answers are stored and fed back as examples. The review_floor setting is a floor on what is worth a human's attention, never a threshold for acting automatically.

Plan limits are enforced by a database trigger rather than a check in the interface, so an over-limit account can still be edited down — downgrading never traps you.

---

Stack

Framework	Next.js 16 (App Router, typed routes), React 19, TypeScript strict
Data	Supabase / PostgreSQL, row-level security on every tenant table
Billing	Stripe — test mode only, no live keys anywhere
Matching	Anthropic (wired); OpenAI and Google are selectable but not implemented
Charts	Recharts
Motion	Native CSS scroll-driven animation (animation-timeline: view()), GSAP for what CSS cannot express
Email	Resend
Capture	Playwright
Tests	Vitest — 56, covering RLS isolation, billing idempotency and matching

Running it

npm install
cp .env.example .env.local     # fill in Supabase, Stripe test keys, Resend
npm run db:migrate             # nine migrations, in order
npm run seed                   # two organisations, ~7,600 price snapshots
npm run dev

npm run typecheck
npm test                       # all suites
npm run test:rls               # isolation only — needs a throwaway Postgres
npm run alerts                 # evaluate the alert engine against seeded history

AGENTS.md is written and re-added by next dev; committing it with your work keeps the tree clean.

Layout

src/app/                 routes — landing, /app dashboard, /architecture, auth
src/app/motion.css       the scroll choreography, and why each range is what it is
src/lib/db/rls.ts        does per request what PostgREST does: sets the JWT claim
                         and the role, so local development has the same boundary
src/lib/data/queries.ts  analytics as SECURITY INVOKER Postgres functions
supabase/migrations/     nine migrations; 0002 and 0007 are the security ones
tests/rls/               adversarial isolation tests
scripts/record/          scripted screen capture for the product clips

Security notes

• Secrets live in environment variables and are never committed.
• Stripe stays in test mode on a dedicated account.
• The crawler visits only the project's own fictional storefronts, rate-limited.
• Third-party marks in public/brands/ and public/providers/ are other companies' trademarks, shown only against capabilities that exist and never presented as a partnership or endorsement. Anything unbuilt is labelled planned in the interface.

Licence

Not currently licensed for reuse.