Work behind this service

AI in production

Three tools in use at Erkan Ulu Schools in Istanbul, and an App Store app built on retrieval.

Erkan Ulu Schools · IstanbulIn production

Agentic OMR answer-sheet reader

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.

  • Teachers photograph sheets instead of typing every answer into the school portal
  • An uncertain read is tried again with a different prompt, then goes to a teacher
  • Results go into the school portal, and the school keeps its own answer sheets
Read the case study
Erkan Ulu Schools · IstanbulLive since May 2026

Ev Çalışması Merkezi

An AI homework platform for a private middle school. Students upload photos or PDFs of handwritten homework; each question is marked correct, wrong or partly correct against the teacher's answer key, and explained.

  • A teacher's override always wins over the AI's verdict
  • After a mistake, students solve a similar AI-generated question by hand on screen
  • Teacher panel with class results, a curriculum-outcome heat map and hand-in tracking
Read the case study
Erkan Ulu Schools · macOSDeployed · v1.0.4

Otocropper

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.

  • Fine-tuned RT-DETRv2 detector for seven kinds of page region, plus Tesseract OCR
  • 97.1% recall and 96.9% precision at IoU ≥ 0.9 on a 238-page book never used in training
  • SwiftUI app with Python, PyTorch and the model inside, signed and notarized
Read the case study
Own app · iOS and AndroidLive on both stores

1in46: AI Dream Interpreter

A voice-first dream journal. It interprets each dream against about 6,300 passages of classical dream scholarship, paints dream art, and can turn a dream into a narrated AI film.

  • React Native / Expo, Supabase with pgvector, Claude and RevenueCat
  • English, Turkish, Arabic (right to left) and Spanish
  • On the App Store since March 2026 and Google Play since September 2026
App Store1in46.com
What you can hire for

What Stratagem builds with AI

Vision models

Detectors and vision-language models trained on your own images, like the YOLO and vision-language pipeline that reads answer sheets.

Marking and review

AI that marks work against an answer key, explains each verdict and lets a teacher override it.

Answers from your own data

Retrieval over a fixed corpus, as in 1in46, which interprets dreams against about 6,300 passages of classical scholarship.

Fine-tuning

When prompting isn't enough: fine-tuned detectors and vision-language models, measured on data they never saw.

Learn more

A person in the loop

Low-confidence results go to a person instead of slipping through. The answer-sheet reader sends them to a review queue.

Serving that fits the job

Models served with vLLM, run offline on the Apple GPU, or called through the Claude and Gemini APIs.

Process

How an AI project runs

  1. Start from real examples

    Collect the photos, scans or documents the system will actually see, including the bad ones: poor light, tilted pages, stray marks.

  2. Keep a held-out test

    Set aside data the model never trains on. Otocropper was judged on a 238-page book that was never used in training.

  3. Build the checks

    Confidence thresholds, retries and a review queue, so uncertain results reach a person.

  4. Ship and keep measuring

    Release to real users, log every model call, and adjust when errors show up.

Questions

Frequently asked questions

Do I need a custom model, or will an API like Claude or GPT do?

Start with an API and good prompts; it's cheaper and faster. Train or fine-tune when the task is narrow and measurable and the API can't reach the accuracy you need, as with finding question regions on scanned pages.

How do you know the model works?

By testing it on data it never saw during training. Otocropper reached 97.1% recall and 96.9% precision at IoU ≥ 0.9 on a held-out 238-page book, and 222 of the 238 pages needed no edits.

What happens when the AI is wrong?

It is built to be checked. The answer-sheet reader sends low-confidence sheets to a review queue, and on the homework platform a teacher's override always wins.

Can a model run without sending data to the cloud?

Yes. Otocropper runs fully offline on the teachers' Macs, with the detector, OCR and Python bundled into one signed app.

What does AI development cost?

It depends on whether you need prompting, retrieval or a trained model. Running costs are broken down in the LoRA fine-tuning cost guide and the RAG cost guide; a build is quoted after a call.

Related reading

Related services

AI

Computer Vision

Vision models tested on data they never saw, like Otocropper.

Learn more AI

Model Training and Fine-Tuning

Custom-trained vision models, measured on a held-out set.

Learn more AI

AI Agents

Tool-using pipelines that check their own work and ask a person when unsure.

Learn more

Have a task you want AI to take over?

Send a few real examples of the work and what a correct result looks like. That is enough to tell whether it needs a prompt, retrieval or a trained model.

Discuss your project See all work

Or call (786) 788-1030

Savas Tutumlu

Who builds it

Savas Tutumlu

Founder & CEO, Stratagem Systems

Started the studio in Houston, TX in 2023 and now works from Istanbul. 30 projects shipped since.