We build AI and automation for lenders, brokers, bookkeeping firms and finance teams: extraction that turns a P&L into editable figures, books that sync themselves, and AI summaries grounded in your own numbers. A language model reads the messy input, ordinary code makes the decision, and a person can correct every figure.
Free call · Fixed-scope quote within 48h
Each one replaced work someone did by hand, in a spreadsheet or an inbox.
Drop in a P&L. An LLM at temperature 0 returns trading income, operating profit, depreciation, interest and non-trading income against a fixed schema, each with a confidence score the broker can override. Code then calculates DSCR and ranks 13 lenders. A 90-minute assessment now takes under ten.
See it in PetranQuickBooks is the spine, bank feeds arrive through Plaid, and receipts attach to transactions by date and amount match. The bookkeeper works a queue with AI-suggested categories and confidence scores, while Stripe Billing handles proration, dunning and failed-payment retries.
See it in TrupennyAn AI column summarises the session, lists takeaways and answers questions grounded in the dashboard’s own data. Two scheduled jobs write a morning and an afternoon summary to a searchable archive. Informational only, no advice logic.
See it in MacroPulseThree finance products where AI or automation does the manual work. Filter the case studies.
An AI-assisted finance broker platform. Upload a client's statements, extract the numbers with an LLM, calculate DSCR, and match the scenario against a live rate-card of lenders, all…
Read the case study
A done-for-you bookkeeping firm needed a client portal that didn't feel like a 1995 web form. We built one, QuickBooks-synced reports, secure document vault, Stripe-billed subscriptions…
Read the case study
A Koyfin-class personal markets dashboard. Forty-plus live metrics across stocks, crypto, commodities, macro, credit, and housing on one surface, with a market-regime read, per-ticker…
Read the case studyLanguage models are very good at reading a scanned P&L, a bank statement or a receipt and turning it into structured figures. They are the wrong place for a lending, eligibility or payment decision, because the same input has to produce the same answer, and the reason has to be explainable to a client, a lender or an auditor.
So in Petran the extraction call runs at temperature 0 against a fixed schema, every figure carries a confidence score, and the broker can edit any of them. The decision that follows is ordinary code: Adjusted EBITDA, DSCR, then a filter over each lender’s thresholds. Same inputs, same shortlist, and the excluded lenders come with the reason they were excluded.
The risk with finance automation is not that it stops. It is that it keeps running with one source stale and nobody notices. In Trupenny every connection (QuickBooks, Plaid bank feeds, Stripe Billing, the document vault) shows its last-synced time and can be resynced in one click.
Revoke any one source and the product drops to a read-only state instead of showing a figure built from half the data. Every number on the dashboard can be traced from the total down to the single transaction and its linked receipt.
Generic AI commentary on markets or accounts is free and worthless. The useful version reads the figures already on your screen and explains them. In MacroPulse the AI column is grounded in the dashboard’s own 40-plus live metrics, so a question like “what is driving the 10-year lower today?” is answered from that data, not from the model’s memory.
Scheduled summaries go to an archive, which turns a daily narrative into a record you can search later. Where advice rules apply, the AI is scoped as informational only, and that boundary is written into the product rather than a disclaimer.
The best first project is rarely “add AI”. It is the step a skilled person repeats all day: re-keying figures from a PDF, chasing receipts by email, copying numbers between systems, writing the same summary every morning. Those steps have clear inputs and outputs, a measurable time cost, and a person who can check the result.
On the scoping call we map that one workflow, decide what the model reads, what code decides and what a person approves, and quote it as a fixed-scope sprint.
Fixed scope, a demo every Friday, and a first clickable version around day 7. The full process is on the home page.
For lenders, brokers, bookkeeping firms and finance teams with a manual workflow that should run itself, inside an existing product or a new one.
2 to 3 wksAdd-onsScheduled AI summaries · client-facing reports · additional lenders or rule sets · monthly tuning retainer
Fixed scope · Fixed price
Demo every Friday
Everything else gets answered on the scoping call.
Mostly reading and routing work: extracting figures from statements and receipts, matching documents to transactions, suggesting categories, and writing summaries of your own data. Petran cut a broker’s credit assessment from about 90 minutes to under ten this way.
No. The model extracts the figures and a person can correct them. The decision itself runs in ordinary, deterministic code, so the same inputs always give the same result and every exclusion has a stated reason.
Good enough for a first draft, never treated as final. Every extracted figure carries a confidence score and stays editable, so low-confidence values get checked before anything is calculated from them.
Yes, through the official APIs. Trupenny uses QuickBooks Online as the source of truth, Plaid for bank feeds and Stripe Billing for subscriptions, each with a visible last-synced time.
No. We use provider API terms that exclude training, send the model only the fields it needs, and set retention explicitly. Card data stays with the payment processor and never enters your stack.
Two to three weeks for an AI or automation layer on an existing workflow, four to six inside a new product. It is a fixed-scope, fixed-price sprint, quoted within 48 hours of the scoping call.
Bring the workflow and a sample document. In 30 minutes we map what the model reads, what the rules decide and what a person approves, and you have a fixed-scope quote within 48 hours.
RaftWorks