An invoice sorter that keeps your clients’ names off the internet

1. 01

   Read the document

   pdf.js pulls the embedded text layer, which is exact and instant. When there isn't one — a phone photo, a flatbed scan — the page is rasterised in the browser and read by tesseract.js, offline. Which route a document took is shown on its row, because it changes how much the row should be trusted.

   No upload directory. The file is read into memory, used, and dropped.

2. 02

   Extract against a schema, not a prompt

   The Zod schema is converted to JSON Schema and passed to the model as a grammar constraint, at temperature 0. Eleven typed fields in or a validation failure out — never plausible JSON with a string where a number belongs.

   Measured: zero schema violations across twenty fixtures.

3. 03

   Score what deserves a human

   Confidence is computed, never asked for. A 4B model asked to rate its own confidence produces a number with no calibration behind it, and it is confidently wrong on exactly the fields that cost money. The score blends whether the value appears verbatim in the source, whether it is well formed, and whether subtotal + tax actually equals total.

   Presented as a review-priority score out of 100. It is not a probability.

4. 04

   Correct, and keep the correction

   Every edit is captured as a labelled example: what the model predicted, what it should have been, and when. That log is the dataset a fine-tune would need, produced as a side effect of ordinary review rather than as a project.

   Downloadable as JSONL. Stored on your disk, or in Atlas if you opt in.

Measured on an i7-1165G7, 4 cores / 8 threads, 16 GB RAM, no GPU, Windows 11. Ollama gemma3:4b Q4_K_M. 20 fixtures, one ambiguous fixture excluded from headline figures and scored leniently. Ground truth authored before the first run and never edited to match model output.

TabPFN wins, by 3%. It is a real margin and it is also a narrow one — who pays late is mostly a property of who the client is, so a well-built per-client mean is a genuinely strong baseline rather than a straw man. The model’s contribution is in the residual: invoice size, timing seasonality and the client’s own variance. Its q10–q90 interval contained the true value in 12 of 17 held-out rows, against an 80% nominal target, so the intervals run a little narrow and are reported that way.

Does my invoice data leave my computer?

On the local path, no. Your browser talks to Ollama at 127.0.0.1 directly, so the extraction request never touches a server and there is no server-side component that could log it. Open the network tab during a local extraction and you will see one destination, and it is your own machine. The app still works with the network disconnected once the model is downloaded.

What happens if I have no model installed?

The runtime banner tells you so plainly rather than pretending. You get a pre-computed dataset — real recorded output from a real run, including its mistakes — so you can exercise review, correction and export without setting anything up. It is labelled as demo data and is never described as live inference.

What is the hosted path, exactly?

An explicit, opt-in alternative for machines with no local model: a 31B model via OpenRouter. It requires one confirmation per session. Client names and account-number-like digit runs are stripped on your device first, requests carry provider.data_collection: deny and require_parameters: true, and if no endpoint satisfying both exists the request fails rather than degrading to something that logs. It is not equivalent to local and the banner does not pretend it is.

How accurate is it, honestly?

66.2% of fields correct across twenty fixtures, and 5.3% of documents fully correct. That is not a system you hand a folder to and forget about — it is a system that does the tedious 60% and shows you exactly which 40% to check. The per-field table on this page says which fields those are, and category scores in the teens.

Why a 4B model and not something bigger?

Because this ran on four CPU cores with no GPU and no credit card. 12B is 8.1 GB and materially slower; 27B would have consumed the remaining disk while remaining unusable. The size curve in the study measures what that constraint actually costs, by running 1B and 4B on identical hardware.

What does the payment prediction actually do?

It forecasts days-to-payment per open invoice with a q10–q90 interval, and it is scored against the naive baselines it has to beat: that client's own historical mean, and the global mean. Only numbers and a hashed client identifier are sent to Prior Labs; the raw client name never leaves your device. Below about twenty historical invoices it is weak, and it says so.

Will it cost me anything?

The local path is $0.00 and there is no checkout anywhere in this app. There is a written $12/month hypothesis in the footer — framed around the chase list rather than extraction — which is a question being tested with the first real user, not a thing you can pay. Payment collection is deliberately absent, partly because Vercel's Hobby terms restrict this deployment to personal, non-commercial use.