Afterscan
Follow-up software for preventive health-scan clinics
A woman in her forties pays a Stockholm clinic 2,750 kr for a preventive scan because her father had a heart attack at 52.
It comes back green, but ApoB and Lp(a), the tests that history calls for, were never on the panel, and the clinic’s app reaches no GP.
Afterscan is follow-up software a scan clinic gives its clients.
In the app, the client uploads her results, confirms the values read and answers four family-history questions. Rules the clinic’s medical lead signed name the missing tests, she books them at a lab the clinic chose, and its doctor signs a letter to her vårdcentral.
Reading each upload into rows of test, value, unit and healthy range takes most of Afterscan’s engineering.
Every clinic app and lab has its own layout, and Lp(a) alone comes in nmol/L or mg/dL. Templates read the known layouts, roughly half of uploads. The rest, every screenshot among them, goes through text recognition, a layout encoder that finds result rows and an extraction model that reads each with a confidence score.
Afterscan’s models read values and nothing more: which tests are missing is decided by the signed rules table, never by a model.
A field read below 0.9 confidence stays blank for the client to type, and every other field waits for her confirmation. Each confirmation leaves a nameless crop of one result row beside its confirmed value, a pair Afterscan owns and fine-tunes on.
Afterscan’s uploads arrive in the evening and at weekends, when clients open their results.
Only the half the templates cannot read goes to GPU inference, so those hours are busy and the week otherwise quiet. Each new clinic adds a burst of its layouts. Health data may not go to an outside model API, so the models run and are fine-tuned on cloud GPU capacity we control in Stockholm.
Afterscan’s reader fails most on screenshots, where any of its three steps today can lose a row.
In Q1 2027 one open-weight vision-language model, fine-tuned on the same pairs, replaces text recognition and both models. Reserved GPU capacity takes the evening and weekend hours from Q3 2027, with fine-tuning after every 5,000 confirmed pairs. Q4 2027 adds embedding search that matches each upload to its nearest past lab layout.
Afterscan S.R.L. was founded in Buenos Aires, Argentina, in September 2022, four years before it could take a single upload.
Those years went into a rules table a medical lead would sign, a letter a vårdcentral would file, and the consent behind every upload.
Afterscan went live in September 2026 on cloud infrastructure in Stockholm, because its clinics’ health data stays in Sweden.
Clinics pay monthly. No clinic or client will appear on this site as an example.
Lucía Ferreyra
Tomás Aguirre
Simone Baptiste