Two systems in use at Erkan Ulu Schools in Istanbul, and three of Stratagem's own apps on the App Store.
A homework platform for a private middle school. Students upload photos of handwritten homework, and AI models mark each question correct, wrong or partly correct against the teacher's answer key.
Reads multiple-choice answer sheets from teachers' phone photos with YOLO and a custom-trained vision-language model served with vLLM, then sends the results to the school portal.
MUST: See, Eat, VisitA city companion for Istanbul, Miami and New York. Its AI concierge answers from the app's own 1,800+ hand-curated places and turns answers into savable plans.
1in46: AI Dream InterpreterA voice-first dream journal. Each dream is interpreted against about 6,300 passages of classical, ancient and Jungian dream literature, retrieved with pgvector on Supabase.
Mawla: Quran, Prayer TimesA daily Islamic companion with prayer times, the Quran and a qibla finder. Its voice assistant answers everyday religious questions with sourced references.
Retrieval-augmented generation (RAG) finds the passages in your own documents or database that match a question, and the model answers from them. 1in46 works this way over about 6,300 passages.
Start with a prompt, add retrieval when answers must come from your sources, and fine-tune only when tests on your data show both fall short. See the RAG implementation cost guide and model fine-tuning.
| Approach | Use it when | In Stratagem's work |
|---|---|---|
| Prompting a hosted model | The task is general and everything the model needs fits in the request. | Homework marking against the teacher's answer key |
| Retrieval (RAG) | Answers must come from your documents or data, and change when they change. | 1in46 (about 6,300 passages in Supabase with pgvector); MUST's concierge, grounded in 1,800+ curated places |
| Fine-tuning | The task is narrow and measurable, and a general model keeps missing it. | The vision-language model that reads answer marks; Otocropper's page-region detector |
An agent splits a job into steps: tools do the fixed work, a model handles the judgment, and checks decide what is saved. The answer-sheet reader runs in four steps.
A YOLO model finds and crops each answer row, allowing for tilt, shadows and creased paper.
A vision-language model served with vLLM reads the marked option, half-erased marks included.
Each read gets a confidence score and a check against the sheet layout. An uncertain read is retried with a different prompt.
Confident results go to the school portal; the rest wait for a teacher instead of being guessed.
Assistants that answer from a defined source and can show where an answer came from. MUST, Mawla and 1in46 each have one.
Retrieval over your documents or database, so answers come from what you have checked.
Answers say where they came from, as in Mawla.
Speech input where typing is awkward: 1in46 records each dream by voice.
MUST's concierge turns an answer into a savable multi-stop plan.
English and Turkish in MUST and Mawla; 1in46 adds Arabic (right to left) and Spanish.
Inside iOS and Android apps built with React Native and Expo, or behind a web chat.
A test set kept apart from the start, with the misses reported.
Uncertain results go to a review queue, and a person can override the model.
The cheapest model that passes your test, and every model call logged.
Results land where the work happens: the answer-sheet reader posts them to the school portal.
Hosted Claude and Gemini APIs, models served with vLLM, or fully offline in an app.
Code, database and hosting set up in your own accounts.
It starts with a free 15-minute call and a written proposal with one fixed price. More on how a project runs and pricing.
Collect the inputs the system will really see, bad photos and tilted pages included.
Set aside data the model never trains on, and agree what counts as correct.
Confidence thresholds, retries and a review queue for uncertain results.
Release to real users, log every model call, and adjust when errors show up.
Whatever fits each step: hosted language models on the homework platform, Claude in MUST and 1in46, YOLO and a custom-trained vision-language model served with vLLM for answer sheets, and FLUX, Kling and Veo for 1in46's dream art and films.
It depends on the design. The homework platform calls hosted model APIs. Otocropper runs fully offline on the teachers' Macs and needs no internet connection.
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.
Usually that is the point. The answer-sheet reader opens from the school portal and sends results back to it, and the homework platform takes its class rosters from the same portal.
Each build is quoted as one fixed price in a written proposal after a call. Running costs depend on the approach; see the RAG cost guide and the LoRA fine-tuning cost guide.
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.
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. He builds every project himself.