AI ENGINEER — ANKARA
most days.
tuesdays are hard.
In production, not just in notebooks — from models and pipelines to the apps people actually use.
I started out studying physics at METU, got distracted by computers, and never quite came back. Getting distracted turned out to be a recurring theme — I’ve just learned where to point it.
I finished a computer engineering degree at Karabük University instead, and kept the habits physics left behind.
Physics taught me to distrust results that look too good. Statistics taught me to prove it.
My first real problems were visual: live video at a simulation company, then deep learning on dental X-rays — where “close enough” isn’t.
TÜBİTAK 2242 · NATIONAL FINALIST · 2022
Graduation project. Deep learning for segmenting, classifying and numbering teeth on panoramic X-rays, trained on a real, expert-labeled clinical dataset. Presented at the finals at TEKNOFEST Giresun.
TÜBİTAK 2209-A · RESEARCH GRANT · 2022
Team of four. Mask R-CNN for tooth detection and FDI numbering, pretrained CNNs for caries classification. Dataset labeled by expert radiologists. Trained on a laptop with a GTX 1650 — which is either impressive or a cry for help.
Today I work on the AI team at Sebit, a Türk Telekom edtech company — building production AI for education and business, plus the unglamorous plumbing that keeps it honest.
Agents that plan, use tools and work together — designed as reusable building blocks, with a human in the loop where it matters.
LangGraph · LangChain · CrewAI · AutoGen · OpenAI Agents SDK · Claude Agent SDK · MCP
Retrieval pipelines that keep models anchored to real content instead of improvising.
LlamaIndex · ChromaDB · Qdrant · pgvector · FAISS · sentence-transformers · rerankers
Gateways, usage tracking and cost control — for hosted models and open-source ones running locally.
LiteLLM · Ollama · LM Studio · vLLM · Langfuse · FastAPI · Docker · Kubernetes · Kafka
Getting reliable data out of messy real-world inputs — scanned pages, forms, handwriting, medical images — and generating new ones when needed.
OpenCV · PyTorch · YOLO · Mask R-CNN · segmentation · PaddleOCR · Tesseract · PyMuPDF · vision-language models · Stable Diffusion
When AI makes the call, something has to check the AI. Consistency testing, cross-model agreement, and anomaly detection to catch outputs that don’t look right.
Ragas · DeepEval · promptfoo · LLM-as-a-judge · scikit-learn
Similarity search, recommendation, time-series and financial forecasting, and predictive models.
OpenAI & open-source embeddings · scikit-learn · XGBoost · statsmodels · Prophet · TensorFlow · Keras · reinforcement learning (Gym)
Turning raw data into something a decision can be made from — and a chart someone will actually read.
pandas · NumPy · matplotlib · seaborn · Power BI · SQL
Education technology · business · finance · healthcare imaging · simulation
LEARNING TECHNOLOGIES AWARDS 2024 · LONDON
Sor O Zaman, with the Sebit team. I built the similarity and recommendation system and the image-processing pipeline, and represented the team in London.
CODiE AWARDS 2025
SEBIT AI Exam, with the Sebit team. I work on its AI side.
ON STAGE · JANUARY 2026
24th Traditional Education Symposium of the Private Schools Association of Turkey, Antalya. Alongside Dr. Pelin Tamcı Tuner (Sebit), moderated by Prof. Dr. Eren Ceylan (Ankara University). On AI as a teacher’s working partner — not a replacement.
Unsupervised anomaly detection that surfaces suspicious decisions for human review.
Problem. A system produces thousands of judgments — some by people, some by models. A few will be wrong, and there are no labels saying which ones.
Approach. Represent each decision and its reasoning as embeddings, then let several independent anomaly detectors vote. Each case gets a severity score, and reviewers decide how far down the list to look.
Status. In active R&D: fewer false alarms, faster at large volumes.
Built from scratch, then hardened in production.
Problem. Scanned forms arrive in large, unpredictable batches, and every page has to be read correctly — exactly once.
Approach. Built the service from scratch with computer vision and an event-driven architecture. Once it was live, I tracked down the failures that only appear under real load, and fixed them at the root instead of patching symptoms.
Status. Running in production. It used to call me at night. Now we’re on speaking terms.
TRY IT
Your picks: 0/3
Three of these 60 points don’t belong. Tap the ones you think are off — then let the model have a go.
The model is a small Isolation Forest running in your browser on synthetic data. It isolates points with random splits; the ones that fall out fastest are the strangest. Numbers are severity scores, 0–100.
Outside work, the same principles run my own projects. The biggest one is an agentic app factory: a pipeline that takes a mobile app from idea to build, run by AI agents and supervised by me over Telegram.
Problem. Building small apps involves a lot of repetitive work: finding ideas, writing specs, coding, store optimization, triaging reviews.
Approach. An orchestrator agent coordinates a coding agent through each stage. Rules keep ideas in scope. Anything risky waits for my approval.
Status. Runs end to end. The apps aren’t published yet — the pipeline is the product.
Work gets done by people, with people. I’d rather be a good teammate than the smartest person in the room — and a good team will outbuild a room full of smart individuals every time.
Keeping things running matters. But I’m at my best building what isn’t there yet — new systems, new capabilities, new ways of working.
A working version in front of real users beats a perfect plan in a slide deck. I’d rather fix what’s actually broken than guess what might be.
Every extra component can break at 3 a.m., and every model call comes with a bill. I add complexity only when the problem demands it, not when the architecture diagram looks lonely.
I let AI do the heavy lifting and keep people on the decisions that carry real risk. Automation earns trust one step at a time.
Models are confident even when they’re wrong. Consistency checks, cross-model agreement and anomaly detection catch mistakes before users do.
If leadership can’t follow the tradeoff, they can’t make a good call. Plain language is part of the engineering. Somewhere along the way, I also became the person who explains AI decisions to people who’d rather not hear the word “embedding.”
I played bass and sang in a band. These days I’m learning drums — the neighbors have been very understanding. Allegedly. I also play guitar, and my playlist jumps from blues and rock to bossa nova, Chopin, Satie and 80s synth without warning.
Logos and characters, for years now — some by hand, some with AI, most somewhere in between.
I run D&D campaigns, leaning into cosmic horror and dark fantasy. My players would describe me as “fair.” Eventually.
Pasta, seared steak, things that make the kitchen smell like a decision was made. Also: Lord of the Rings, Star Wars, noir, and dark comedy.
Updated: October 2026
Got a messy AI problem? I like those.
Email is the fastest way to reach me. I read everything and reply to most things.
Earlier writing: From Blackboards to Bots (LinkedIn, 2024) → Certificates →