Software Engineer — Generative AI
I work across text, image, audio, and video, and wire them into a single pipeline. Before AI, I spent 10+ years building games and interactive work.
Selected work[ about ]
I'm a software engineer working across AI, games, and interactive design. I build multimodal generation pipelines myself: LLM orchestration, image and video generation, TTS, and music. Then I ship that AI into interactive products.
[ selected work ]
One prompt in, a narrated video out.
Give it one prompt and it plans the film, generates the storyboard images, records multi-voice narration with word-aligned captions, animates the stills, drops in a music bed, and composes the final cut. Text, image, audio, and video, all in one pipeline. Live now, with real billing.
Visit montager.app→
A panel of LLMs that agree on an answer.
A deployed Turkish platform that rates games for child safety. At its core is a multi-model consensus engine: it runs OpenAI, Claude, and Gemini in parallel, reconciles the three verdicts by median score, flags how much they disagree, and stays conservative when they do. Structured-output validation and cost-capped workers keep it in check. No media generation here, just text.
Visit dijital-fener.web.app→
Drop a pin, get a historical panorama.
Turns a point on the map into an AI-generated 360° historical panorama. It stitches Google Street View tiles into one equirectangular image, checks the scene with a vision model, then generates the panorama through a queued image model that calls back securely when it's done. Credits are atomic and refunded if a job fails. Image and text.
Walkthrough on request
A multi-provider pipeline studio.
Wraps fal, KIE, and Gemini behind one engine and runs them through an approval-gated, retrying state machine that passes context from one step to the next. Its brander sub-pipeline pins a brand's identity (palette, fonts, logo), then generates on-brand creative around it, with background segmentation and editable text overlays.
Walkthrough on request
Fine-tuning SDXL from the weights up.
An SDXL fine-tuning stack with a full training loop, from noise scheduling through weight export. The trained weights then run on self-hosted SDXL and FLUX inference, behind a queue that watches GPU memory. Everywhere else on this site I call hosted models; here I train one. Images only.
Walkthrough on request
A durable engine that grows video, segment by segment.
A generation engine written in Go. A Postgres-backed task queue feeds a worker pool that picks up where it left off after a restart and retries failed work on its own. On top of it, a recursive cascade grows a tree of generated segments — text, images, video, and audio — each stage waiting on the ones before it, every new segment seeded from the last frame of the one before.
Walkthrough on request
[ let's make something ]
emretokerler@gmail.com