What we store, and for how long
| What | How | How long |
|---|---|---|
| The statement PDF and its rows | In memory for the request only | Dropped when the response is sent |
| A PDF's password | In memory, to open the file | Never stored or logged |
| Async job results | Encrypted | 1 hour after the job ends, then deleted |
| A statement kept so you can ask questions again (opt-in) | Encrypted | What you asked for (retain_minutes), at most 24 hours |
| Request log | Counts, timings, bank name, error codes and model cost. No transactions, names or amounts | Used for your usage charts |
| Webhook delivery log | Event ids, status and counts. No statement data | 30 days |
| Account | Email, password hashed with argon2, linked Google or GitHub id | Until you delete the account |
| API keys | Hash only. The key is shown once | Until revoked |
| A file you attach to a bad-parse report (opt-in) | Only if you choose to attach it | Until the parser is fixed |
Pages that need a language model are sent to our model provider, Anthropic, under its API data terms. Pages read from the layout, and scans read by free OCR, never leave our server. An account or a single API key can switch AI off completely.
How a number becomes trustworthy
Every balance is re-checked
Each row's balance must equal the previous balance plus credit minus debit, to the kobo, oldest first or newest first. Breaks are listed with their page and row.
Model readings are matched against the page
When a language model reads a page, every value it returns is matched word for word against the page's own text. A number that is not printed on the page is caught, not trusted.
Repairs are reported, never silent
Months stapled into one PDF are split into sections, each balanced on its own. Rows a bank printed twice are removed and listed. A brought-forward line is not counted as money in.
A grade your code can act on
Every response carries quality high, medium or low with plain-word reasons. Auto-accept high; route the rest to a person.
How we protect the service and your customers
Questions in English never show the statement to the AI
The model turns your question into a query. Our code runs the query on the checked rows, and you get the data plus the exact query, so the answer can be repeated without the model.
AI can be switched off
An account or a single key can turn AI off. Every page is then read on our servers only, scans included, through free OCR, and English questions are off. GraphQL still works.
Webhooks are signed and cannot be aimed inward
Events are signed with HMAC-SHA256 and a timestamp, carry ids and counts but never statement data, and are only sent to https addresses that resolve to public IPs, checked again before every send.
Keys and access
Live keys are hashed at rest and revocable, and can be limited to certain IP addresses or an expiry date. Sessions are httpOnly cookies with an origin check. Sign-up, login, reset and parsing are rate limited.
Heavy work is isolated
Every parse runs in a separate worker process, so a huge or broken file fails only its own request and your other calls stay fast.
What we do not claim
We do not hold SOC 2 or ISO certifications today and we do not publish an uptime figure yet. You must have your end user's consent to process their statement.
Read the Privacy Policy and Terms of Service for the legal wording. To report a security problem or ask a question, email support@sumproof.dev. These facts are also available as markdown.