Bugrahan D.

I build production AI systems that run 24/7 for real customers: computer vision pipelines, generative image/video pipelines, and LLM/VLM pipelines. Seven years in, currently a Senior ML Engineer at an ad-tech company serving thousands of businesses, plus an ongoing Upwork contract for a US computer-vision SaaS. What I do best: COMPUTER VISION (detection, tracking, evaluation) ● Object detection and tracking in video, from sports ball-tracking and pose estimation to logo detection (YOLO, PyTorch, OpenCV) ● Evaluation harnesses with per-class precision/recall gates and leak-safe holdout sets, so you know the real number, not the model-card number ● Model optimization and deployment on GPU cloud (GCP, RunPod) and on edge/mobile hardware GENERATIVE IMAGE & VIDEO ● AI product photoshoot pipeline: products placed into generated scenes with consistent lighting and pixel-accurate masking (Gemini image models, fal.ai, custom inpainting/blending) ● UGC-style product video generation from a single photo (Seedance, Veo3, AI voiceover) ● Stable Diffusion / ControlNet / LoRA fine-tuning, virtual try-on, and a SAM + harmonization blending engine that fixes the "pasted-on" look LLM & VLM PIPELINES ● VLM document-extraction pipeline (Gemini/Vertex AI, OpenAI, Groq) with JSON-schema enforced outputs: invalid responses cut from 36% to 0%, extracted output up 45% with zero silent drops ● LLM ad-copy generation and VLM creative scoring in production, consistent across multiple model providers ● Prompt systems, structured outputs (Pydantic), tracing and evals (Langfuse) Stack: Python, PyTorch, OpenCV, FastAPI, Docker, GCP (Cloud Run, Vertex AI), AWS, RunPod, PostgreSQL, TypeScript/Next.js when the job needs a UI. MSc Computer Engineering. Fluent English, clear written communication, documented handovers. UTC+3, overlapping with both EU and US mornings. Tell me about your pipeline and I'll tell you how I'd build it, and how I'd prove it works.