Alex P.

This project developed a RAG-enhanced voice assistant combining automated HR support and psycho-linguistic screening, analyzing speech for psychological distress markers while ensuring data privacy (GDPR/HIPAA). Core workflows were built with LangGraph for dynamic routing between HR responses and screenings, integrated with Pydantic-AI for output validation. Security features included sandbox environments and incident response playbooks linked to Azure App Insights. A custom RAG architecture used LangChain with Llama-3, integrated human review with LangSmith traces for transparency, and vector embedding storage in PostgreSQL. Prompt management governed production prompts with iterative experiments to boost accuracy. Models such as BERT, GPT, and Llama were fine-tuned on clinical datasets using LoRA. Solutions were wrapped as FastAPI services, deployed in containerized Azure environments with Kubernetes orchestration and automated CI/CD pipelines via GitHub Actions.

An application for finding resources for caretakers and people with low social status. The Applications main feature was a chatbot that could recommend organizations for getting resources such as food, clothing, specify the illegibility criteria and apply for grants. The target users of this platform are caretakers and guardians of foster children. For the organization there is an internal version of the application that allows creating tickets for grant requests and processing them Environment Python, Pandas, Hugging Face, ONNX Runtime, Langchain, Scikit-Learn, XGBoost, PostgreSQL, Docker, Terraform, Azure(AI Studio, Document Intelligence, AI search, Language, CosmosDB, OpenAI, Container Registry, Container Apps, DataBricks, Synapse, Machine Learning, etc.), Git, Azure DevOps. Led architecture, CI/CD (Azure DevOps/Terraform), Azure ML workspace. Built LangChain RAG agents, PII scrubbing, BERT fine-tuning for QA/grant prediction. 5x faster ticket handling via recommendation system.

Spearheaded development and optimization of a mission-critical application for processing industrial sensor data. Focused on predictive analytics to forecast equipment remaining useful life (RUL), estimate failure probability, detect operational anomalies in real-time, etc. Led cross-team collaboration aligning ML, MLOps, engineering, and business goals. Developed ETL pipelines for sensor data with Apache Spark. Designed data lake architecture with governance, schema enforcement, and quality gates. Performed data cleaning, feature engineering, and created training datasets. Managed training environments with Azure ML, MLflow, Tensorboard. Integrated approval gates enforcing performance and explainability. Improved anomaly detection by 12% F1 score. Applied reliability theory for lifecycle modeling. Automated model selection/tuning with MLBox, KerasTuner. Reduced RUL RMSE from 300 to 10 via neural ensembles. Built custom SDK for training and scheduling. Used RL to optimize maintenance