A multi-tenant SaaS point-of-sale platform for small/medium retail, pharmacy, restaurant, salon and service businesses, with built-in FIFO/WAC inventory costing, tiered pricing, multi-branch stock transfers, cash-drawer/bank reconciliation, and PayHere subscription billing.
DynaPOS runs one company's multiple businesses across multiple branches with strictly tenant-isolated data. It models accounting-correct inventory (FIFO stock layers with per-sale cost snapshots, WAC dispatch, decimal units of measure), tiered/quantity-break pricing, inter-branch stock transfers that preserve cost lineage, end-of-day cash-drawer-to-bank reconciliation, and PayHere subscription billing — all from a single codebase serving 9 industry templates, with a Playwright + Cucumber BDD suite running in CI on ephemeral Postgres.
Role & Context
Lead developer in a small team — owned end-to-end architecture and the majority of implementation: the Prisma schema and multi-tenant model, RBAC, accounting primitives (FIFO stock layers, WAC dispatch, decimal UoM), PayHere webhook integration, and the Playwright/Cucumber E2E harness. Scope spans v1.0 (auth, onboarding, dynamic per-business-type product schema, POS, inventory) through v9.0 (cash drawer, bank ledger, deposit batching): 18 versioned migrations and 9 industry templates from one codebase.
Problem
Small/mid-size businesses in Sri Lanka had to choose between generic global POS products that don't model LKR / 18% VAT / PayHere, per-vertical point solutions that lock a merchant into one business type, or spreadsheets that lose audit trails on stock and cash. None handled the combination DynaPOS targets: one company running multiple businesses across branches with strict tenant isolation, accounting-correct FIFO/WAC inventory across transfers, tiered pricing, end-of-day cash-to-bank reconciliation, and local payment rails.
Approach & Architecture
Loading diagram…
Auth.js-gated Next.js App Router resolves a request-cached tenant context, Zod-validates, checks RBAC, then runs FIFO/WAC + pricing inside prisma.$transaction; PayHere webhooks are MD5-verified and idempotent on lastPaymentId.
Best Practices Followed
Defence-in-depth tenant isolation: tenant context resolved once per request via React cache(); every Prisma where scopes by companyId/businessId; a partial unique index enforces one open cash drawer per cashier at the DB layer
Accounting integrity by construction: all stock/cash mutations wrap in prisma.$transaction; SaleItemConsumption snapshots per-layer unitCost at sale time so historical P&L is immutable even after a FIFO<->WAC switch
Decimal precision discipline: money as DECIMAL(12,2), fractional quantities as DECIMAL(18,4) for UoM conversions, with explicit rounding at every write boundary
Schema-validated boundaries: every API entry point Zod-parses before any side effect; webhook payloads use the same pattern
BDD-style E2E in CI: ~210 Cucumber feature files driven by Playwright cover sales, returns, transfers, billing, RBAC and 5 business types, with an idempotent seed for deterministic fixtures
Targeted composite indexes for hot paths: StockLayer(productId, branchId, receivedAt) for the FIFO walk; Sale(branchId, createdAt) for daily reports
Challenges & Resolution
ChallengeInter-branch stock transfers had to preserve original layer costs without breaking FIFO accounting — naive transfers would either lose cost lineage or double-count COGS.
FixModelled StockTransferItem.costBreakdown as JSON snapshotting the (layerId, qty, unitCost) tuples consumed from source layers; the receiving branch spawns mirror StockLayer rows at the original unitCost, so a unit's cost basis survives any number of inter-branch hops.
ChallengePayHere webhooks can be re-delivered (retries, manual replays), and duplicate processing would double-charge subscriptions or corrupt billing state.
FixVerified the MD5 notification signature first, Zod-parsed the payload, then keyed idempotency on lastPaymentId per subscription so a duplicate payment_id short-circuits to a no-op; the design distinguishes initial-charge vs renewal correctly.
Outcomes
Shipped to production on Vercel, serving real merchants across multiple business types (multi-tenant)
18 versioned, forward-only Prisma migrations from v1.0 Foundation to v9.0 Cash & Banks, including a v7.6 backfill that spawned opening-balance StockLayer rows for every (product, branch) so FIFO/WAC could light up over historical inventory
9 industry templates (mobile shop, pharmacy, bookshop, grocery, salon, restaurant, hardware, general retail, services) from a single codebase via dynamic product field definitions
~210 Cucumber/Playwright E2E scenarios running in CI on an ephemeral Postgres, with a Serenity/JS living-documentation report uploaded per run