Generative AI + computer vision. Virtual try-on that runs on a laptop GPU. FLUX.2 Klein redraws only the clothes, guided by clothes parsing and pose estimation. It estimates your size from a photo and your height, then draws how a size you can buy would really fit. Every photo and result passes three independent safety checks. About 15–25 s per look on an RTX 4060.

PyTorch Diffusers SegFormer MediaPipe CLIP OpenCV FastAPI React

Real-time perception for connected vehicles (graduation project). Custom-trained YOLOv8n (mAP@0.5 0.67), OpenVINO optimisation (39 FPS on CPU vs 30 with PyTorch), centroid and Kalman tracking, monocular distance estimation, lane detection and ROS nodes. The live demo runs detection in your browser.

YOLOv8 OpenVINO ONNX OpenCV Kalman filter ROS

LLM / RAG, Arabic + English. Answers questions about your documents with citations, across languages. Hybrid search (BM25 + multilingual embeddings) with re-ranking reaches Recall@5 1.00 and MRR 0.967 on a 30-question bilingual test set. Works with Gemini, any OpenAI-compatible model, or no key.

RAG multilingual-e5 FAISS BM25 cross-encoder FastAPI transformers.js

NLP. Scores a CV against a job description with SBERT and TF-IDF, lists matched and missing skills, and flags biased wording. ROC AUC 1.00 and 95% accuracy on 20 labelled pairs; 22 tests, CI.

SBERT scikit-learn FastAPI pytest Docker GitHub Actions

MLOps. Categorises retail transactions (TF-IDF → SVD → clustering → classifier), packaged with MLflow tracking, a FastAPI service, drift checks and CI that trains and tests. 97.1% accuracy on descriptions the model never saw, with an honest note on what pseudo-labels can and can't prove.

scikit-learn MLflow FastAPI Docker GitHub Actions

Explainable classical ML. Seven models compared; the deployed logistic regression reaches 91.3% accuracy and AUC 0.935 on a held-out set, explains each prediction, and ships with a model card.

pandas scikit-learn PCA model card

LLM product. A CEFR language-learning platform where Gemini checks learners' answers and explains mistakes, with a local fallback when the model is unavailable.

Gemini Next.js React

A portfolio you can browse with hand gestures (MediaPipe hand landmarks in the browser) or by asking an assistant.

MediaPipe TensorFlow.js Next.js FastAPI

At Poseidon IT I work as a full-stack engineer on web, mobile and API work. That is why my models come with a product around them. Some examples:

• 🏥 Virtual clinic · ▶ live demo: video visits, payments and prescriptions (React, Express, MongoDB)
• 📊 Arabic sales operations: an RTL-first workspace for field sales teams (React, TypeScript, PostgreSQL)
• 📱 Climate Edge: a Flutter + Firebase document workflow app

I take freelance projects such as:

• Object detection & tracking for retail, security or traffic video
• RAG chatbots over your documents, in Arabic and English
• Document & CV matching, semantic search, text classification
• Virtual try-on and image editing with self-hosted generative models
• Notebook → production: turning a model into a tested, deployed API

📧 nardymichelle2003@gmail.com · 💼 LinkedIn

_{📈 Open the interactive version to hover any day for its exact contribution count.}

• Now: Full Stack Engineer at Poseidon IT: frontend, mobile, backend, APIs and databases.
• Before: computer vision, machine learning, data analysis, Flutter development and technical operations.
• Graduation project: cooperative perception for connected vehicles (YOLO, OpenVINO, lane detection, Kalman tracking, ROS).