generative ai / llm systems
LLM assistants, retrieval workflows, and practical AI
I am a Generative AI Engineer at Nokia, working on LLM assistants, evaluation loops, retrieval workflows, and AI-powered internal tools.
Ahmet Barış Emre Durak
I like building AI systems that are useful after the demo: measurable, observable, and steady enough for real workflows.
featured project
Super-Resolution for Digital Elevation Model Data
A deep learning graduation project comparing SRGAN and ESRGAN architectures for enhancing Digital Elevation Model data — reconstructing higher-resolution terrain from lower-resolution elevation inputs.
Low-resolution Digital Elevation Model data loses terrain detail, weakening downstream terrain analysis, mapping, and geospatial workflows — and high-resolution data is costly to collect in data-poor regions.
Trained and compared SRGAN and ESRGAN generator/discriminator networks on SRTM elevation data from Türkiye and Austria, combining adversarial, reconstruction, total variation, SSIM, and MS-SSIM losses.
A comparative study: both architectures improved on traditional interpolation, with ESRGAN producing sharper, more structurally accurate terrain — plus early evidence that models can transfer across regions with similar terrain.
notes
Latest Notes
From Contrastive Learning to World Models
A technical roadmap through visual self-supervised learning, from negative-pair contrastive objectives to explicit embedding regularization, latent prediction, and V-JEPA-style world models.
Notes on Reliable RAG Systems
A short placeholder note for future writing about retrieval quality, latency, and LLM assistants.
Super-Resolution for Digital Elevation Model Data
A graduation project on enhancing DEM resolution with SRGAN and ESRGAN, using terrain data from Turkiye and Austria.