About NitsorHosted · by invitation
Nitsor is a 3D annotation tool for CT and MRI scans. Every label it saves carries an author, a reviewer and a date, and every release carries a proof someone outside your company can check. The scans themselves stay in your own storage.
The system of evidence for AI-assisted inspection.
That sentence is short enough to be useless on its own, so here it is in plain words.
Inspection is deciding whether a thing is sound: a casting, a weld, a prostate. AI-assisted means a model now does part of that deciding. Evidence is the part everybody skipped. When a model helps make a call, somebody has to be able to show, later, what the call rested on: which scan, which version of the labels, whose judgement, and which model.
That record is the product. The drawing tools exist so that the record has something to record.
A small team, working on this full time. The hosted instance runs at app.nitsor.com and signing in to it is by invitation, so every account so far started with a conversation.
We are not going to put a wall of stock photographs on this page, and we do not have a customer list to show you. What we have instead is public: the datasets we build and test on, with their permanent identifiers, and a list of what works today that we keep honest even when it costs us a call.
The posts in Notes from the build are written by the people who wrote the code that week. There is no marketing team between them and you.
From inspection work, and from the same afternoon happening over and over: a model flags something, an expert disagrees, and nobody can reconstruct which version of the labels the model was trained on. The tools for drawing on medical and industrial scans are good. The tools for remembering what was drawn, and by whom, and under which rules, were a shared drive.
Custody, not supply. A labelling vendor takes your scans and sends finished work back. We leave the scans where they are and keep the record instead. You keep custody of the pixels; we keep the account of what was done to them, and you can take that account away whenever you like.
A label is a change with a name on it. Not a file in a folder. One event per object, with the person or model that drew it, the person who accepted it, the version it replaced, and the time. That is what makes "which labels trained this model" a question with an answer.
Automation only as far as you can prove. There is one dial for how much the model does on its own, and it turns only as far as a measurement covers. A high confidence score is not that measurement; models are confidently wrong exactly where it matters. Above the bar, predictions skip human review only where the measurement says they may, and the measurement is taken on cases the model never saw.
Most of these are promises about what we will never charge for or take away. They are worth writing down because the usual practice in this market is the opposite, and because a promise you can read is easier to hold us to than one you have to infer from a price list.
The list lives on one page and nowhere else, so it cannot drift into three versions of itself. Ten sentences, dated, with a fingerprint you can recompute in your own browser: the deployment promises.
The short version. Reading is free forever. Read-only people are free forever. Nothing is capped. We never charge per person or per label. Provenance is never behind a paywall. Export is complete and unconditional, whether or not you are still paying us. If a page of ours ever says something softer than that, the dated one wins.
Nitsor is source-available. You can read the code that holds your record, and run it, under a licence that stops someone reselling it as their own service. Prices are not set, and the pricing page says so rather than inventing figures.
Every picture on this site is a slice of a published scan, with the dataset's own labels drawn over it. We show these because a company with no customers to name should show its working instead of a logo wall.
Both datasets are free to download under a licence that only asks for credit, so anything we measure, you can measure too. The counts in our posts were taken from these files and the notes say when.
A very large computed-tomography scan of a historic aircraft section, published by Fraunhofer EZRT with instance labels. It is the closest public stand-in for a casting or a composite part, because it has hundreds of separate objects in one volume.
A repeat-scan study from The Cancer Imaging Archive, with the study's own DICOM SEG labels. It is small, it is real, and eleven of its thirty slices carry a label, which is the shape that makes routed review worth doing.
The label colours are ours, chosen so that every instance sits at the same perceived brightness and no label shouts louder than another. The instance names are ours. The labels are the datasets'. Any moving picture of glass or machinery on this site is generated illustration and is captioned as such wherever it appears.
Thirty minutes on your own scans, on our hosted instance, no slides. If we are not a fit, we say so on the call.