ML / Computer Vision Engineer. Obsessed with soldering and programming gadgets out of literal garbage. Building things that actually work.
Based in Wrocław. I fine-tune vision models, build the datasets behind them, and ship AI tools for European B2B companies: GDPR-compliant, production-grade, no bullshit architectures.
Currently deep into defence tech, computer vision, model fine-tuning, datasets & evaluation, AI agents, MCP tools and embedded systems - because apparently building one thing at a time is not an option.
Mom of two. Mass-producing code and humans with equal enthusiasm.
| Project | About | |
|---|---|---|
| 🪟 | nadya-vitrazh | Interactive stained-glass installation - YOLO person detection + MediaPipe pose + state machine. |
| 🐭 | dormouse | Ukrainian text optimizer for LLMs - 73% token savings, 150% quality. pip install dormouse-ua |
| 🔧 | Vibecode-Cleaner-Fartrun | Rust-powered code scanner with 29 MCP tools. Not another AI reviewing AI. |
| 🪢 | knot-detector | Real-time knot-tying detection: MediaPipe HandLandmarker + heuristic. In-browser demo + Python prototype. |
| 🗓 | mcp-ukrainian-calendar | MCP server for Ukrainian pagan calendar - holidays, deities, folk wisdom. |
WAWTech 2026 - speaker, 26-27 November, EXPO XXI, Warsaw.
Як я зекономила 50-73% на GPT API, бо в мене комп зі смітника - про
dormouse-ua, написаний на Phenom II 2009 року.
class Daria:
location = "Wrocław, PL"
role = "ML / CV Engineer"
kids = 2
coffee = float("inf")
focus = ["Defence Tech", "Computer Vision", "Model Fine-tuning",
"Datasets & Evaluation", "AI Agents", "MCP Tools"]
learning = ["edge export: ONNX -> Hailo", "MLOps"]
passion = ["Linux Kernel", "Raspberry Pi", "ESP32"]
hero = "Linus Torvalds" # unironic
addiction = "sci-fi" # mass effect > sleep
spirit_animal = "Sheldon Cooper" # BAZINGA!
@property
def available_hours(self):
return 24 - self.kids * 3 # optimistic estimate🛡 Defence tech: computer vision (details stay private)
🎯 Fine-tuning detection and retrieval models (YOLO, SigLIP + ArcFace/SupCon)
🔎 Visual product search: recall@1 57 -> 73 after fine-tuning
🧪 Synthetic training data: Blender + domain randomization, mixed synth/real
🗂 Dataset pipelines: harvesting, CLIP dedup, Grounding DINO pre-labeling, human review
💎 Photo -> castable 3D jewelry: SAM 2, Depth Anything V2, photometric stereo, Blender
🤖 AI agents and MCP tools in production
🔌 Embedded: Raspberry Pi, ESP32, sensors




