AI Engineer and solo SaaS founder. I build software that ships, mostly around NLP and transformer models, with an MSc in Information Technologies behind it. Right now I run FersmithAI, a learning platform for people moving into AI/ML roles. It's live, and subscriptions are open.
Built and launched FersmithAI solo, from the first commit to paid subscriptions. It's a learning platform that turns dense technical material into daily study plans for people moving into AI/ML while working full time: 28 skills, 93 audio episodes from my own TTS pipeline, 16 hands-on labs, 902 flashcards, 720 interview questions and a 140-question assessment. It runs on Next.js 16, Supabase and Paddle.
Completed a remote data analyst internship, gaining hands-on experience in data analysis workflows and reporting.
Cabin crew on international flights, and still flying while I build FersmithAI. The job runs on safety procedures, calm decisions under time pressure and clear communication in English and Turkish.
Translated academic and creative texts (EN<->TR) with precision and deadline discipline. Built a strong foundation in text quality, structure, and linguistic nuance, the same instincts NLP work runs on.
GPA: 3.14/4.00
I work as an AI-assisted engineer: I design the system, split it into steps, and review and test what gets built. Claude Code, Codex and GitHub Copilot write much of the code, and I check it. Everything below is something I've used in my own projects.
FersmithAI: Personalized AI/ML Learning Platform Live
A subscription learning platform for people who hold a full-time job and have 1-2 hours a day to move into AI/ML. Every morning it asks two questions, how many minutes you have and what you're working toward, then builds that day's plan to fit the answer.
The library covers 28 skills across 7 categories, each with 3 podcast episodes (~20 min), a video, a worked code example, 20 flashcards, and 20 interview questions. Around that sit 16 hands-on labs (Docker, Git, MLflow, LangGraph, kubectl and more), a 125-card glossary, and 9 paper breakdowns. 902 flashcards and 720 interview questions in total, with all 93 audio episodes produced through my own TTS pipeline. An optional 140-question assessment scores you skill by skill and reorders your queue, and it never locks content away.
Most people stitch this together from three separate subscriptions: one for courses, one for practice, one for interview prep. Here it's one, and all of it is AI/ML. Built solo on Next.js 16, React 19, TypeScript, Tailwind CSS 4, Supabase (Postgres, Auth, Storage), Paddle for billing, and Vercel for hosting.
SkyAnalyst-AI: Cabin Operations Intelligence Framework
A modular NLP framework that transforms unstructured flight cabin reports into structured operational insights using RAG and abstractive summarization. Built on ChromaDB vector storage and all-MiniLM-L6-v2 embeddings for semantic retrieval across 2,500+ synthetic incident reports, so search runs on meaning instead of keywords. Summarization is powered by DistilBART-CNN, delivering 4x faster inference than standard BART with negligible quality loss. Includes an exploratory analytics dashboard with sentiment intensity scoring and statistical visualizations across 5 incident categories and 22 flight routes.
Children’s Tale Summarizer: Master’s Thesis, NLP Pipeline & Flask App
At the time, Turkish had very few language models trained on creative writing, and the available ones (T5, BERT, GPT variants) pulled sentences out of a text instead of writing a real summary. For my master’s thesis I built an end-to-end pipeline that writes abstractive summaries of Turkish children’s tales and illustrates them.
I tested more than 15 transformer models on 13 tales, then fine-tuned the top three with detailed hyperparameter optimization. That reached up to 2× ROUGE-L (0.25 to 0.52) and +3 points of BERTScore F1.
The Flask app takes a PDF or TXT file, summarizes the tale in three parts (introduction, development, conclusion), translates the summary into English and draws an illustration for each part with Stable Diffusion. Authors can get original illustrations for a book without drawing them, and 90% of users found the visuals matched the story.
