ProdBusiness

ProdBusiness is an enterprise B2B ERP for wholesalers, distributors and industrial suppliers. One explainable engine covers pricing, procurement, stock and finance. A navigable workspace prototype and a 480-table PostgreSQL model that enforces correctness and tenant isolation in the database are built; the FastAPI backend is planned.

Many ERPs keep price, invoice and stock truth separately for each channel. Staff orders, portals, EDI and field agents can then disagree. Multi-company books, warehouse-scoped stock and B2B exceptions (credit holds, price mismatches, missing receipts) end up spread across modules and spreadsheets. AI features are often bolted on without evidence, approvals or an audit trail. ProdBusiness's own predecessor showed these limits: one legal company, globally keyed stock and a scheduler tied to the web process.

ProdBusiness is designed around one authoritative calculation and execution core shared by every channel. PostgreSQL itself enforces balanced journals, immutable posted ledgers, sealed price decisions, single-reservation stock and tenant/company row-level security. The UI is role-first and explainable: work queues, entity workspaces, a price waterfall showing applied and skipped rules, plain-language invoice-match explanations and decision previews before any commit. AI is limited to proposals: it shows evidence and a before/after diff, and a human approves before any action, which is recorded as an action receipt.

• Role-first workspaces

  Implemented

  Twelve module workspaces (Sell, Buy, Stock & Ship, Finance and more) across 90 routes. They include a command palette, persona switcher, record timeline drawer and English/French UI.

• Procurement suite

  Partial

  Requisitions, RFQs with a bid comparison matrix, purchase orders, proforma invoices and landed cost, all built as working mock screens.

• Explainable invoice matching

  Partial

  Price, quantity, tax, duplicate and missing-receipt mismatches are explained in plain language, with status and next step.

• Price waterfall

  Partial

  Each pricing step shows whether it was applied or skipped, why, and which rule version produced it.

• Warehouse and shipping

  Partial

  Receiving and two-stage receipt, bin positions, scan-based picks, blind counts, shipments, carriers and load planning.

• Finance workspace

  Partial

  Chart of accounts, journals, open items, a bank reconciliation split view, period close and e-invoicing queues.

• Database-enforced invariants

  Implemented

  Balanced journals, immutable ledgers, sealed pricing, idempotent commands and no double stock reservation, verified by 30 passing checks on embedded PostgreSQL.

• Tenant and company isolation

  Implemented

  Forced row-level security with restricted runtime roles. A missing tenant context fails closed.

• AI proposal drawer

  Partial

  AI suggestions appear as uncommitted proposals with purpose, evidence, a before/after diff, uncertainty and an approval boundary.

• FastAPI domain services and workers

  Planned

  A versioned /api/v1 command API, a durable outbox, integration workers and Redis-assisted caching.

• EDI, carrier and e-invoice integrations

  Planned

  EDI partner onboarding, DHL/freight booking, structured e-invoices and bank feeds through certified adapters.

Two layers are built today: a Next.js App Router workspace prototype running on an in-memory mock store, and an executable PostgreSQL reference model validated in embedded Postgres. The target runtime is a domain-modular FastAPI monolith. It writes canonical state to PostgreSQL, publishes through a durable outbox to separate domain and integration workers, and uses Redis only for reconstructable caches and rate limits. AI tools may only call allowlisted domain commands through approval-gated proposals.

frontend

Next.js (App Router, Turbopack)ReactTypeScriptlucide-reactIBM Plex

backend

FastAPI (planned)Python (planned)SQLAlchemy / Alembic (planned)

data

PostgreSQLRow-level securityRedis (planned)

ai

Policy-bound proposal and receipt modelProvider-agnostic (planned)

infra

PGlite validation harnessNode.js toolingS3-compatible storage (planned)

integrations

EDI / Peppol (planned)DHL and freight (planned)ZUGFeRD / XRechnung (planned)DATEV and FEC exports (planned)

• Reviewable AI proposals

  A dedicated drawer shows an AI suggestion's purpose, supporting evidence, before/after diff and uncertainty. It is clearly marked as not committed until a human approves.

• Governed AI data model

  Thirteen schema tables cover model policies with spend limits, allowlisted tools and grants, budgeted runs, proposals, evidence citations, action receipts and safety evaluations.

• Planned grounded assistance

  Roadmap items: document extraction with citations, quote and price explanation, bid summaries, forecasting and exception triage. Autonomy is tiered from read-only up to approval-required financial actions.

1. 01

   Ledger correctness inside the database

   Deferred constraint triggers reject unbalanced journals at commit and block mutation of posted lines. Accounting truth does not depend on application code.

2. 02

   Concurrency-safe stock reservation

   Scoped writers and guards ensure the last unit cannot be reserved twice and a retried posting batch cannot deduct stock twice.

3. 03

   Multi-tenant, multi-company isolation

   474 row-level security policies with restricted roles, composite scoped foreign keys and fail-closed empty context, all checked by automated tests.

4. 04

   Keeping 480 tables coherent

   A machine-readable schema model, a feature-to-entity map and Node validators keep table, column and FK counts consistent with the docs and 32 ER diagrams.

• Explainability first

  Price waterfalls, plain-language match explanations and decision previews show users why something happens before they commit.

• Correctness by construction

  Money, stock, idempotency and isolation rules live in PostgreSQL guards rather than only in services.

• AI that proposes, never silently acts

  Every AI suggestion carries evidence and a diff and requires approval. Executed actions leave a permanent receipt.

• Falsifiable product goals

  Each candidate differentiator has a measurable pass/fail evidence gate instead of a marketing claim.

The source repository is private. A live walkthrough is available on request.