Aleksandr

Cross-platform crypto wallet for secure multi-asset management, real-time market tracking, and encrypted transactions. Features biometric auth, multi-wallet support, transaction history, offline-first mode, encrypted storage, JWT auth, and blockchain monitoring, with multilingual UX and CD pipelines. Responsibilities: Worked closely with business/dev teams. Implemented a multi-agent system for complex crypto queries and personalized interactions. Built LangChain agents for wallet operations (balances, transfers) and ReWOO-style LangGraph agents orchestrating multiple tools. Optimized agent workflows with ReAct/self-refinement for better planning and reliability. Used LangSmith and OpenRouter for LLM observability and debugging. Integrated 3rd‑party crypto/finance APIs, built RAG with PostgreSQL + pgvector, Redis-based address book, FastAPI/Pydantic ML microservices, CI/CD, and MCP Server, aligning AI features with security and UX.

RAG‑enhanced voice assistant for HR that both answers policy questions and performs preliminary psycho‑linguistic screening for well‑being, combining RAG with privacy‑first analysis of distress markers.​ Responsibilities: Led sessions with HR/domain experts to refine goals, risks, and vision; drove the business proposal phase and made AI design transparent for non‑technical stakeholders. Designed a LangGraph policy enforcement layer to enforce GDPR/HIPAA on medical/HR data and delivered a production MVP with sensitive screening under strict privacy controls. Built a custom RAG stack with LangChain + Llama‑3, human‑in‑the‑loop escalation using LangSmith traces, and vector storage in PostgreSQL + pgvector. Fine‑tuned transformer models with LoRA on curated clinical datasets; evaluated models and RAG for fairness and ML metrics. Optimized retriever via chunking/embedding experiments. Wrapped solutions as FastAPI services, containerized with Docker, deployed on Azure, and set up Jenkins

Led development of an industrial predictive maintenance platform for sensor data, focusing on RUL forecasting, failure probability, and real-time anomaly detection.​ Collaborated in a 20+ person cross-functional team. Built Spark ETL, data cleaning, feature engineering, and Parquet training sets. Set up SageMaker + MLflow + TensorBoard with drift and artifact tracking and approval gates (RMSE/F1 + explainability) before prod. Upgraded LOF anomalies with Isolation Forests (+12% F1). Introduced LSTM+1D‑CNN ensembles cutting RUL RMSE from 300 to 10, aligned with CNN–LSTM best practices.​ Developed an SDK for ensemble training and SageMaker job scheduling, end-to-end MLOps on MLflow + Kubernetes, and IG-based explainability. Mentored junior DS and conducted technical interviews. Stack: Python, TensorFlow/Keras, TensorBoard, MLflow, FastAPI, Spark, Docker, K8s, GitLab CI/CD, Kafka, AWS (EC2, SageMaker, ECR).