Alexandru Paicu

Revolutionary RAG system achieving 97% storage reduction through on-demand embedding computation. Implements LEANN (Lightweight Embedding & Neural Network) with HNSW graph backend that stores only navigable small-world graph structure instead of dense vectors. Features real-time PDF/text ingestion with adaptive chunking (256-token blocks), Qwen3-0.6B LLM for response generation, and all-MiniLM-L6-v2 embeddings. Optimized for CPU deployment on Hugging Face Free Tier (2 vCPU, 16GB RAM) with 2-5s search latency. Graph-traversal architecture recomputes embeddings during search, eliminating index bloat while maintaining semantic accuracy. Supports GPU acceleration (10x faster on T4/A10G) with zero code changes.

Binary classification model predicting sleep disorder risk using NHANES survey data (2005-2016 cycles). Engineered features from lifestyle and health variables including BMI, exercise, diet, sleep duration, and cardiovascular metrics. Implemented XGBoost as baseline model with Optuna hyperparameter tuning. Integrated SHAP explainability for global and local feature importance analysis and counterfactual recommendations. Deployed with Streamlit UI for real-time risk prediction and personalized health recommendations.

Autonomous code generation system with self-healing capabilities using dual-model architecture (3B/7B Qwen2.5-Coder). Implements LangGraph state machine orchestrating ReWOO planning framework (Planner-Worker-Solver). Features advanced RAG 2.0 pipeline with RAPTOR hierarchical indexing, HyDE hypothetical document generation, MultiQuery expansion, and FlashRank cross-encoder reranking. Includes iterative self-correction loop: Generate → Execute → Reflect → Fix (max 3 iterations). Supports document upload for context-aware code generation. Built with async Gradio streaming interface and ChromaDB vector store with BGE-base-en-v1.5 embeddings.

End-to-end housing-price regression pipeline built with production ML engineering best practices: time-aware splits, robust preprocessing & feature engineering, XGBoost with Optuna hyperparameter tuning, MLflow experiment tracking and containerized deployment. Includes an interactive Gradio UI and a REST API, plus CI/CD and AWS-ready task definitions.

Developed classification models using Random Forest, SVM, and Neural Networks on UCI heart disease dataset. Achieved 87% accuracy through feature engineering and hyperparameter optimization. Implemented cross-validation and ROC analysis for model validation.

Created regression models predicting vehicle prices using ensemble methods. Processed 15K+ vehicle records with feature engineering on categorical and numerical data. Reduced prediction error by 23% through advanced feature selection techniques.

Built time-series regression models for heavy equipment auction price prediction. Achieved RMSE reduction of 15% compared to baseline linear models. Implemented model retraining pipeline for continuous learning.

Developed CNN using TensorFlow and Transfer Learning with pre-trained ResNet models. Achieved 85% classification accuracy across 120 dog breeds. Created web interface for real-time image classification.

Developed Python scripts for processing Gaia DR3 and SDSS astronomical datasets. Created data visualization tools for stellar motion analysis and cosmic event pattern recognition. Implemented statistical models for anomaly detection in large-scale astronomical data.

Engineered a serverless AI code generation tool using Anthropic Claude models via AWS Bedrock, integrated with FastAPI for real-time API access.

Developed an AI-powered image generation tool using foundation models via AWS Bedrock, enabling rapid creation of high-quality visuals.

Built a serverless text summarization tool using Anthropic Claude models via AWS Bedrock, integrated with FastAPI for API-driven access.