Interest
Split recurring borrowing costs from one-off facility and establishment charges. Recurring interest stays in the servicing calculation; one-off interest is added back.
?Is any part of finance costs a one-off establishment or break fee?
Built for Australian mortgage brokers
The income figure — and the paper trail behind it.
Fin-XTract reads Australian financial statements, applies your add-back rules across current and prior year, then puts Claude over every figure to reconcile it against the page it was taken from.
No document is ever stored — processed in memory, discarded the moment the run completes.
0
documents retained after a run
2
independent passes over every figure
CY + PY
extracted in a single pass
< 90s
median time to a results pack
The results pack
Every figure carries its own evidence
Why this exists
An income figure is only as good as the argument behind it.
Every lender wants the same thing — a defensible number. Getting there means reading a set of financial statements the way the accountant actually wrote them, not the way a template expects them: finding the add-backs, holding current and prior year side by side, and being able to point at the page each figure came from.
Generic OCR gives you text. A spreadsheet gives you arithmetic. Neither gives you the argument.
Twelve accountant styles
The same P&L line is called four different things across four firms. Alias matching handles the vocabulary, not just the layout.
Scanned pages that lie
A vision pass reads the page image and reconciles it against extracted text, so a smudged 8 never quietly becomes a 3.
Judgement, not automation
Interest, one-offs and disposals are surfaced with the evidence and a suggested question — the decision stays with the broker.
How it works
Five stages, each one inspectable
The pipeline is deliberately split so an accuracy problem can be traced to a single stage — a classifier miss is not the same failure as an alias miss, and neither is an arithmetic error.
- 01
Ingest & validate
PDF, DOCX, JPG or PNG up to 10 MB and 15 pages. MIME type, size and page count are checked before a single byte is processed, and nothing is written to disk.
PDF · DOCX · JPG · PNG10 MB15 pagesNo persistence - 02
Classify
A discrete classifier stage identifies the accountant's house style, the entity type — company, trust, partnership, sole trader — and which pages actually carry the financials, with a confidence flag on each.
House styleEntity typePage selectionConfidence flag - 03
Extract & calculate
Attachment A aliases are matched across current and prior year columns. Add-backs and subtractions are applied by rule, and the Total is derived deterministically so it can be re-run and reproduced.
CY + PY columnsAlias matchingAdd-back logicReproducible total - 04
Verify with Claude
Claude Sonnet, via Bedrock, re-reads profit, depreciation and totals against both the page image and the extracted text, returning strict JSON with a confidence score. Token spend is logged per call.
Vision + textStrict JSONConfidence scoreCost logged - 05
Investigate & export
Attachment B items are rendered with their source snippet and a suggested question for the accountant. The signed-off pack exports to PDF and counts against the subscription quota.
Source snippetsAccountant letterPDF exportQuota counted
What we extract
Attachment A is the arithmetic.
Attachment B is the judgement.
The ruleset is explicit and versioned. Nothing is inferred from a model's general knowledge of accounting — the aliases below are matched, and anything that cannot be matched is reported as absent rather than guessed.
Attachment A
15 fields · CY + PY- Net profit before tax
- NPBT · Operating profit · Profit before income tax
- Depreciation
- Depreciation expense · Depn & amortisation
- Amortisation
- Amortisation of intangibles
- Interest expense
- Finance costs · Interest paid · Borrowing costs
- Superannuation — director
- Director super · Superannuation contributions
- Director remuneration
- Directors' fees · Management salaries
- Trust distributions
- Distributions received · Beneficiary entitlement
- Rent — related party
- Rent paid to related entity
- Motor vehicle expenses
- MV expenses · Vehicle running costs
- Lease & hire purchase
- Finance lease charges · HP interest
- Donations
- Donations & gifts
- Non-recurring income
- Other income (one-off) · Abnormal items
- Profit on asset disposal
- Gain/(loss) on sale of assets
- Franking credits
- Imputation credits
- Retained earnings movement
- Movement in retained profits
Attachment B
Requires broker judgementOne-off income & expenses
Abnormal or non-recurring items that distort the year — insurance recoveries, legal settlements, COVID-era grants, restructuring costs.
?Which items in other income are not expected to recur next year?
Asset disposals
Book gains and losses on sale of plant, vehicles or property, which are removed because they do not represent trading income.
?Was the gain on disposal of the vehicle a one-off, and is a replacement planned?
Security & data handling
The document leaves no trace
Client financial statements are among the most sensitive documents a broker handles. Fin-XTract never persists one — files are held in memory for the duration of a run and discarded the moment the results pack is produced.
No document storage
Uploads are processed in memory and never written to a bucket, a database or a log. There is no retrieval endpoint because there is nothing to retrieve.
JWT with refresh rotation
Short-lived access tokens, rotating refresh tokens, bcrypt-hashed passwords, mandatory email verification and self-service password reset.
Payments never touch us
Stripe Checkout and the Stripe customer portal are hosted. No card number, expiry or CVC ever reaches a Fin-XTract server; webhooks are signature-verified.
Scoped multi-user accounts
Several brokers can work under one subscription, each with their own login and their own usage trail, while quota is enforced at the account level.
Per-user audit trail
Every extraction and every export is attributed to a named user with a timestamp, so a brokerage can reconstruct who produced which figure.
Logged AI spend
Each Claude call records its token count and cost against the run, so the unit economics of the AI layer stay visible rather than becoming a surprise.
Pricing
Priced on exports, not on effort
Run as many extractions as you like while you are working a file. You are only charged when you export a signed-off results pack — that is the artefact that goes to the lender.
Solo
For an individual broker running their own files.
A$89per month
- 1 user
- 25 exported packs per month
- Unlimited extraction runs
- Attachment A + B in full
- Claude verification on every run
- PDF results pack
- Email support
Brokerage
Most chosenFor a team sharing one subscription and one usage trail.
A$349per month
- Up to 5 users
- 120 exported packs per month
- Everything in Solo
- Team management & invitations
- Per-user usage reporting
- Shared extraction history
- Priority support
Enterprise
For aggregators and lenders with their own ruleset.
Customannual
- Unlimited users
- Custom export quota
- Your own Attachment A ruleset
- SSO and provisioning
- Dedicated environment
- Named support contact
- Volume pricing
All prices in AUD and exclusive of GST. Cancel or change plan at any time through the Stripe customer portal.
Questions
The ones brokers actually ask
They are held in memory for the length of the run and discarded when the results pack is produced. There is no storage bucket, no document table and no retrieval endpoint — if you need the file again, you upload it again.