fedorsolovev

Led data platform engineering and hands-on implementation of enterprise data and AI infrastructure for urban mobility, powering both operational analytics and production AI agent workflows. Data Platform & Production AI Systems: - Architected and operated the production mobility data platform: ~180 active Airflow DAGs across 38 business and technical domains (rides, payments, telematics, parking, relocations), Airbyte CDC (18 connectors), ClickHouse, PostgreSQL, and OpenMetadata. - Engineered incremental ELT pipelines using watermarks, upserts, and lineage tracking, ensuring strict data synchronization between operational reporting and AI agent contexts. - Implemented comprehensive data quality controls: dbt models, custom SQL integrity assertions, freshness monitoring, and automated reconciliation with real-time alerting via Telegram and Bitrix. - Built the data foundation for AnalystGraph (compound NL-to-SQL agent): metadata indexing from OpenMetadata, lineage graph traversal, BGE context reranking, and schema-guarded SQL generation (certified across 12 business test suites). - Built geospatial analytics microservices and aggregated data marts with FastAPI and GeoPandas over ClickHouse and PostgreSQL (78 Moscow districts, 500+ operational zones). - Deployed a scalable GPU inference cluster on NVIDIA A100 (vLLM, Qwen 35B, embedding/reranking services) in Kubernetes, exposing governed data context to agents via FastMCP. Enterprise Knowledge Infrastructure (GraphKnowledge Platform): - Architected a multi-modal ingestion pipeline for the corporate knowledge base GraphKnowledge: transcribing video meetings (DION STT / protocols), Confluence documentation, Jira tickets, GitLab repositories, OpenMetadata catalogs, and system logs. - Built an automated cross-repository reconciliation system matching live Git code against Confluence/Jira technical documentation to flag documentation drift and outdated architecture specs. - Implemented automated technical documentation generation pipelines powered by local LLMs grounded on dependency graphs and vector code snippets. - Engineered the data pipeline for a multi-agent project

Designed and productionized enterprise data and knowledge infrastructure for industrial AI diagnostics and predictive maintenance platforms across chemical manufacturing plants. Designed an ontological knowledge graph in Neo4j (2M+ nodes covering equipment hierarchy, maintenance logs, and operational telemetry) serving as shared semantic memory and context for AI agents. Engineered hybrid retrieval pipelines: GNN-assisted subgraph extraction, Milvus vector search, BM25 lexical search, Refrag context restructuring, and vLLM inference. Built high-throughput batch and streaming industrial data pipelines using Kafka, Airflow, Spark, ClickHouse, PostgreSQL, InfluxDB, and MinIO. Optimized ClickHouse and PostgreSQL for high-frequency IoT workloads: partitioning, sharding, replication, and data retention policies. Implemented a governed data-access layer for AI agents: asynchronous FastAPI microservices and FastMCP servers with typed schemas, input validation, and access boundaries across DWH, graph, and technical document repositories. Established end-to-end provenance: linking raw Kafka event streams and sensor telemetry to knowledge graph entities and engineering documentation for fully explainable AI diagnostics. Operated production environments on Kubernetes/k3s with GitOps pipelines (FluxCD, ArgoCD, GitLab CI), HashiCorp Vault, and observability stacks (Prometheus, Grafana, Loki). Suc

Led analytics architecture and data engineering delivery: scalable Airflow ETL pipelines, Metabase reporting, and transition toward CDC (Debezium) and distributed processing via Apache Spark. Architected digital ticketing data integrations (Telegram Web Apps) connecting payment acquiring, fiscal services, and transit system APIs. Established data quality frameworks and automated operational reporting for transit operators, municipal agencies, and public-sector stakeholders.