FinTech / Accounting automation / Applied AI
InvoiceFlow
AI accounting automation for SMEs
Small businesses that run several companies receive invoices and receipts by email, by phone photo and on paper, then re-key every amount, tax and currency into a spreadsheet at month end. InvoiceFlow centralizes those documents per company, extracts the accounting fields with a multimodal model, converts foreign currencies to CAD, flags duplicates and marks anything uncertain for human review.
- 16 default expense and revenue categories per company
- 4 conditions that route a document to human review
- FR · EN bilingual interface, GST/QST/HST handled natively
- Status
- In production
- Model
- B2B
- Market
- Canada (Québec first)
- Website
- invoicempower.com
The problem
Bookkeeping for a group of small companies is a document problem before it is an accounting problem: supplier invoices, receipts and revenue confirmations arrive in three formats and two languages, some in foreign currencies, some twice, some with three invoices in one PDF. Each one has to be read, its GST/QST split out, its category chosen and its total converted before anyone can trust the month’s numbers.
The solution
A multi-company platform that takes documents in — upload, device camera or a forwarding address per company — and turns each one into a reviewed transaction: a multimodal model extracts the fields, a deterministic layer repairs dates, converts currencies at the invoice-date rate and detects duplicates, and every document below the confidence threshold is held in a review queue until a human confirms it. Dashboards and CSV exports close the loop. InvoiceFlow is a separate product from FinSync, my ERP-to-CCH Tagetik integration platform: no shared code, a different problem. Next on the roadmap: direct exports to accounting packages — today the platform exports CSV.
What the platform does
- Document intake Multiple upload of PDF, JPG, PNG and HEIC files with browser-side compression, capture from the device camera, and email intake through a connected mailbox with one expense and one revenue address per company — receipts written in the body of an email are processed too.
- AI field extraction Supplier, invoice number, dates, subtotal, GST/QST/HST, total, currency, payment method, line items and a suggested category from the company’s own list, each with a confidence score; a deeper re-analysis can be run per document or in batch.
- Human validation Any document under the confidence threshold, suspected duplicate, multi-invoice file or doubtful date is marked for review; every field and line item stays editable, and the raw model response is kept with the document for audit.
- Foreign currencies Currency detected from symbols and supplier country, conversion to CAD at the rate of the invoice date from public sources with a fallback, rate and CAD total stored separately and editable by hand.
- Duplicates and multi-invoice files Duplicate detection at analysis time — automatic removal only when invoice number, supplier and amount all match an older record — plus an on-demand deterministic scan across a selection, and split on request when several invoices share one file.
- Multi-company and roles Each company has its own currency, fiscal month, tax rates, categories and logo; accounts are approved by an administrator and carry a role — administrator, manager, accountant with write-only access, or read-only viewer — enforced in the interface and in the database.
- Dashboards and exports Monthly revenue, expenses, net balance and review backlog, a six-month trend, yearly revenue versus expenses, category breakdown, and a CSV export that carries currency, exchange rate and CAD total for every transaction.
Architecture
Server-rendered web application whose server functions sit behind an authentication middleware; PostgreSQL with row-level security scoped to each company on every table and on file storage; access-restricted object storage organized by company, year, month and document type; PDF compression in the browser before upload to respect the server limit; document processing and mailbox synchronization called from the server side only, with a scheduled endpoint for automatic synchronization.
Who uses it
Owners and administrative staff of small businesses that manage one or several companies, and their accountant through a dedicated role that can upload, correct and categorize without deleting.
Markets
Canada, Québec first: GST and QST rates by default, optional HST, CAD as the pivot currency, French and English interface.
Where AI fits — and where it does not
Documents are processed server-side by a large multimodal model called through a cloud AI gateway: the whole document — PDF, image or email text — is sent for analysis together with the company’s category list, and the model returns a constrained JSON extraction. Everything after that is deterministic code: date repair, currency conversion, duplicate detection and the review decision. A human validates every extraction below the confidence threshold, the raw model response is kept with each document, and the on-demand duplicate scan uses no AI at all. No retrieval layer, no fine-tuning: the only context the model receives is the document and the category names.
My role
Founder and architect: product vision, functional specification, platform architecture and technical direction — from the extraction pipeline to the data model and access rules; my team at SOFT M POWER CANADA built it and operates it under my direction.