Level digitizes thousands of handwritten forms with DocuPipe

Level is an educational nonprofit that brings self-paced educational and workforce development programs to people who are incarcerated. They mail printed curriculum guides into correctional facilities, and each guide ends with a Completion Form the learner fills out by hand, in pen, and mails back. It may look antiquated from the outside, but it is the most reliable delivery method there is: people in prison do not have access to the open internet.

A single returned form packs in a lot that Level needs:

All of it has to reach Level's database before a completion certificate can be issued or a learner's progress tracked. That meant staff and volunteers physically reading and keying every field of every form by hand - slow, expensive, and impossible to scale as the program grew.

The shortfall didn't show up as a queue anyone could point at - it showed up as forms nobody ever got to. In the year before DocuPipe, Level transcribed 47% of what came back. More than half of everything learners wrote and mailed in was never processed at all.

Level went looking for something that could read handwriting and bubble answers automatically. They searched the web and worked through options from large tech providers and open-source models, and none of it looked easy, affordable and consistent at the same time. Then they found DocuPipe.

Level set the whole thing up themselves, with no onboarding team: one extraction schema per guide, each mapping that guide's form to the exact database fields they were already using for manual entry. Their system sends an API call with a scan of the returned form and gets the structured transcription back on a webhook - free-text boxes as text, quiz answers and ratings as clean enumerated values. A new guide takes a new schema and little else, and a form comes back transcribed in under a minute for cents.

The reading is the hard part. Learners write in cursive and block print, sometimes in all caps, sometimes in light pencil, often outside the lines. Bubbles get X'd or slashed instead of filled. Whole sections come back blank. The paper itself arrives folded and torn from the mail, and some forms are photocopies with institutional printing stamped over them.

That is exactly what DocuPipe is built for. Level tested accuracy against their own staff, and after refining the schemas found that DocuPipe read the forms more accurately than the people who had been keying them.

Level's staff now hand-processes 5% of returned forms. The other 95% are transcribed automatically and land in the database within a minute of the form being received, with no headcount added. They have also gone back and processed thousands of forms from previous years that had been sitting untranscribed - data from learners that nobody at Level had ever been able to read.

More than 5,000 forms a year, at roughly 20 structured fields each, adds up to over a hundred thousand data points a year that used to be typed in by hand. That time goes to learners now instead of paperwork.

Level built and owns a schema for each of its dozen-plus completion forms, and updates them itself in plain language whenever a form changes. No code, no engineering project.

Reliable, immediate transcription changed what Level could offer, not just how fast they could process what they already had.

Learners had been asking for more personalized programming for years. Later in 2026, Level is launching an offline AI career mentor built on the same foundation: people in prison document their skills, work experience, education and location on a paper form, DocuPipe transcribes it, and Level's engine returns career options, training recommendations and local employment opportunities specific to that person. None of it works without fast, accurate, affordable handwriting transcription on the front end.