A safety-first mobile product for discovering people nearby and turning that proximity into an actual conversation.
Arda Erdoğan.
AI × SOFTWARE ENGINEER
Ideas are good.
Built is better.
I work with AI, write software, and turn things I’m curious about into things you can use — products people open, and experiments that only ever had to answer a question.
See the workImplementation Engineer @ Neara
Master’s in AI @ KTU
Less talk.
More shipped.
From a local business to a mobile product. Different problems, different scales, the same urge to get something working and put it in front of people.
A website for a family gelateria in Gbely, Slovakia, where the daily flavours have to be both the easiest thing to update and the easiest thing to read.
A website for a Turkish natural stone business that has to convince an architect and a homeowner at the same time.
Still figuring
things out.
The other side of building: understanding what’s underneath. These are questions I couldn’t answer by reading, so I wrote the code instead.
Image ↔ language
An image captioning model built from scratch — a ResNet18 encoder and a transformer decoder that writes a caption one token at a time — paired with CLIP so the same system also runs in reverse, from a sentence to an image.
PyTorch · ResNet18 + transformer decoder · Flickr8kLearning to generate
Can a classifier learn from data that was never real? A conditional WGAN-GP generated the fruit, and across a fifteen-experiment grid, training on its fakes alone stayed within one to two percent of training on photographs.
PyTorch · conditional WGAN-GP · 15-experiment grid
The shape of writing
Handwriting modelled two ways on the same EMNIST letters — as a picture, by a DCGAN, and as a movement, by an LSTM predicting pen strokes — to find out what each one actually captures.
PyTorch · DCGAN vs LSTM-MDN · EMNIST
More than one process.
Most weeks are split between shipping software that has to work for real people and reading about methods that might not work at all yet.
I like that split. The research side keeps me asking why something works; the engineering side keeps me honest about whether anyone can actually use it. Everything on this site comes out of that back-and-forth — products people opened, and experiments that only ever had to answer a question.
Curiosity Research Engineering Product
Stay in
the loop↗
Follow the code, read the thinking, or just see what happens next.