Onur Kirman · Software Engineer, AI
I work across agentic workflows, model evaluation, and production infrastructure — combining hands-on engineering with a research background in computer vision and generative models.
Production AI at Deel · Research background in computer vision and generative models
Selected work
Two examples from production engineering and research. The production details are intentionally sanitized; the architecture and lessons are real.
Production AI · sanitized case study
A tool-grounded reasoning workflow for legal-entity migration across heterogeneous schemas, designed with evaluation and human review as first-class stages.
~80% less human labor · 4× faster processing
Research · IEEE TUAC 2021
A line-segmentation method for handwritten text recognition, benchmarked on the IAM dataset and built around debuggable separation of concerns.
Peer-reviewed publication · public implementation
Recent writing
Writing is where I examine agent architecture, evaluation, frontier-model behavior, and the connections between current systems and earlier ML research.
The latest npm worm is a reminder that release age, lockfiles, deterministic builds, and install-time defenses all belong in a practical supply-chain security baseline.
Tool-grounded reasoning gets you most of the way. The last mile is evaluation, and nobody tells you that part up front.
A short, practical mental model for DDPMs — from a thesis I actually had to defend.
Experience
My work has moved from robotics and computer vision into data platforms and production AI—with the same focus on systems that hold up outside a demo.
Earlier work · research, vision, and robotics
Engineering approach
Uncertainty, observability, and human review are product requirements—not cleanup work after a model is connected.
Ground decisions in source data and bounded tools so generated reasoning stays connected to real state.
Measure intermediate and final outputs so failures become attributable, repeatable, and useful.
Route uncertainty and policy-sensitive decisions to people before they cross an operational boundary.
About
I started with robotic systems and computer-vision research, where separating failure modes was part of the scientific process. Data engineering added scale, ownership, and operational constraints. Production AI brought those lessons together.
Today I am especially interested in how frontier models behave once they are surrounded by tools, policies, evaluation, and real consequences. I use writing to work through those questions in public.
Education
Start a conversation
Especially around agent architecture, evaluation, frontier-model behavior, and the realities of shipping applied ML.
onur@kirman.dev