An offline Mac app that turns scanned question-bank PDFs into one image per question, named by its printed number and filed by unit and test.
Reads answer sheets from phone photos and scans. YOLO finds the answer rows, a custom-trained open-source vision-language model served with vLLM reads the marked options, and strict checks decide what is saved.
A short written answer first: prompting, retrieval or training, with the reasons and the running costs.
Collect, label and split your data. Draft labels pass a verifier and a person before they reach training.
Fine-tuning on a CUDA workstation, more than one checkpoint, and thresholds calibrated on the validation split.
A held-out test set and a strict metric, reported with the misses.
Served with vLLM, run on-device with PyTorch, or packaged inside an app.
New versions are run on real documents before they ship. Otocropper 1.0.2 was checked against 20 of the school's PDFs, 4,813 pages.
Measure how a general model does on your data before training anything.
Label real examples and keep a test set apart from the start.
Train, then choose thresholds on validation data only.
Score the frozen model on the test set, publish the result, and release.
Prompting first. Retrieval when the model needs your documents. Fine-tuning when the output format or the visual task is narrow and the general model keeps missing. The PEFT methods guide compares the options.
In practice the GPU time is often the smaller part; preparing data and evaluating take longer. Current GPU prices and worked examples are in the LoRA fine-tuning cost guide.
Open models you can own: a vision-language model served with vLLM for reading answer marks, and RT-DETRv2 for finding page regions. Licences matter: Otocropper avoided AGPL code on purpose.
Results are reported on a held-out set the model never saw. Otocropper's numbers are on its case study page.
Describe the task and how a general model does on it today. If prompting or retrieval will do, you will hear that first.
Or call (786) 788-1030
Who builds it
Founder & CEO, Stratagem Systems
Started the studio in Houston, TX in 2023 and now works from Istanbul. 30 projects shipped since.