AI Agent Development

An agent earns its place when a job breaks into steps that tools can do and a model can check. The answer-sheet reader Stratagem built for Erkan Ulu Schools works that way: detect, read, verify and save, with a second read and a teacher review when a result is uncertain.

Discuss your project Read the agentic OMR case study

Agentic pipeline in production
Verify and retry built in
Human review for low confidence

Work behind this service

Agents and assistants in use

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

Own app · iOSLive on the App Store

MUST: See, Eat, Visit

A city companion for Istanbul, Miami and New York: a daily brief with weather and transit status, a map of 1,800+ hand-curated places, an events feed, and an AI concierge whose answers turn into savable multi-stop plans.

Expo / React Native, TypeScript, MapLibre and Claude
The concierge answers from the app's own curated places
English and Turkish, re-themed per city

App Store

What you can hire for

What an agent project includes

Tool-using pipelines

Each step done by the tool that suits it: a detector finds, a model reads, a rule checks, and your system stores the result.

Self-checks

Results are checked against what is expected before they are saved; an uncertain read is tried again with a different prompt.

A person in the loop

Anything the agent isn't sure about goes to a review queue, not into your records.

Grounded answers

Assistants that answer from your own data, the way MUST's concierge draws on the app's curated places.

Changes you can trace

Every change an automated job makes is recorded with a timestamp, as in the Google Ads engine's change journal.

Connected to your systems

Results go where the work happens: the answer-sheet reader submits its results back to the school portal.

Process

How an agent project runs

Split the job

Write down the steps a careful person would take, and which of them a tool can do.
Tools before prompts

Detection, lookups and validation as plain code; the model only does the steps that need judgment.
Checks and a review queue

Decide what the agent may save on its own and what goes to a person.
Measure and tighten

Look at what reaches the review queue, then tighten the checks where errors appear.

Questions

Frequently asked questions

What makes a pipeline 'agentic'?

The job is split into separate tools, each with its own check, and the pipeline acts on what the checks find. The answer-sheet reader finds rows with YOLO, reads them with a vision-language model and checks the result against the sheet layout before saving. An uncertain read is tried again with a different prompt, and if it is still uncertain, a teacher decides.

Can an agent act without supervision?

Only where a mistake is cheap. In the systems on this page, uncertain results go to a person, and on the homework platform a teacher's verdict overrides the AI's.

Which models do you use?

Whatever fits the step: YOLO and a custom-trained vision-language model served with vLLM for answer sheets, Claude for the assistants in MUST and 1in46, and Gemini and Claude on the homework platform.

Will it work with our existing software?

That is usually the point. The answer-sheet reader is a web app opened from the school portal for a given student and test, and it submits the results back to the portal.

Have a multi-step job an agent could do?

Describe the steps a person takes today and where mistakes would be costly.

Discuss your project See all work

Or call (786) 788-1030

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.

AI Agent Development

Agents and assistants in use

Agentic OMR answer-sheet reader

MUST: See, Eat, Visit

What an agent project includes

Tool-using pipelines

Self-checks

A person in the loop

Grounded answers

Changes you can trace

Connected to your systems

How an agent project runs

Split the job

Tools before prompts

Checks and a review queue

Measure and tighten

Frequently asked questions

What makes a pipeline 'agentic'?

Can an agent act without supervision?

Which models do you use?

Will it work with our existing software?

Related reading

Related services

Chat and Voice Assistants

AI Development

System Integration

Have a multi-step job an agent could do?