Dilshod Utaganov
A decade of building the data platforms companies run on — ingestion that never silently drops a row, warehouses that stay fast as they grow, and machine learning wired straight into the business.
Products and companies I have built data for
what I do
Ten years across data engineering, analytics, machine learning and platform architecture — from the first ingested event to the dashboard an executive opens on Monday morning.
Batch and streaming ingestion, OLTP to OLAP replication, API and event collection, Parquet processing at scale. Idempotent loads, real backfills, and alerting that fires before the business notices.
Medallion layering for clarity, Data Vault 2.0 where history and auditability matter. Models that survive a re-org, a new source system, and the next five years of questions.
Churn prediction, advertising payback forecasting, and AI agents that take routine work off the team. Models leave the notebook and become scheduled, monitored, retrainable services.
Rewriting the pipelines nobody wants to open: cutting warehouse spend, collapsing multi-hour DAGs into minutes, replacing brittle cron chains with real orchestration.
selected work
Retail, fintech, insurance, food delivery, research and gaming — different domains, the same job: make the numbers arrive on time and be right.
architectures
Seven architectures across three clouds. Different tools, one rule: the layer a person queries should never be the layer the data landed in.
Bronze, silver and gold layers built on Delta tables and transformed with PySpark on Databricks. Azure Data Lake Storage holds the landing zone, Azure Data Factory drives ingestion and change capture from operational systems, and Power BI reads semantic models served straight off the gold layer — so the business never queries raw data and every figure traces back to a versioned Delta table.
A raw vault of hubs, links and satellites loaded from operational systems, a business vault holding calculated logic, and dimensional marts on top — all modelled in dbt and orchestrated by Airflow. History stays fully auditable, and onboarding a new source system becomes an additive change instead of a rewrite.
Third-party APIs and Kafka event streams landed into BigQuery as Dagster assets, giving every table visible lineage and freshness. dbt owns the silver and gold transformations, and Metabase is pointed exclusively at the gold models, so analysts always read curated tables rather than raw ingestion output.
GA4 and Firestore events streamed through Kafka into ClickHouse, then folded into materialized views built on AggregatingMergeTree and ReplacingMergeTree engines, so aggregates stay correct under late-arriving and duplicated events. Dashboards read pre-aggregated tables instead of scanning raw event history.
Kinesis streams feeding Lambda functions written in Python, Go and Node.js, with SQS and SNS for buffering and fan-out and Step Functions coordinating multi-stage jobs. Glue catalogues the lake, DynamoDB serves low-latency lookups, and Athena queries Iceberg tables directly — real-time processing without a cluster to babysit.
Dataflow handling streaming and batch transforms, Dataproc running Spark for heavy reprocessing, Cloud Storage as the lake, Dataform managing SQL transformations inside BigQuery, and Looker as the semantic layer the business actually works from.
LLM agents wired into internal tools and the warehouse: answering data questions in plain language, drafting and reviewing SQL, summarising anomalies and keeping documentation current. Routine requests stop queuing behind the analytics team, and engineers get their review cycles back.
toolbox
Each lane is a layer of the platform, running on its own belt. Hover a lane to stop it and read what is on it.
64 tools · 11 layers
before the pipelines
National scholarships and olympiad awards in physics — the habit of solving hard problems started well before the first data warehouse.
Awarded to senior students for academic excellence and contribution to university research, in national competition.
NRNU MEPhIGranted to a single student for academic results and scientific activity.
NRNU MEPhINational award for academic excellence and research participation.
NRNU MEPhIGranted to a single student for merit in study and university life.
NRNU MEPhIWinning the national stage granted university admission without entrance exams.
Republic of UzbekistanNational stage, physics.
Republic of UzbekistanBuilding a data platform, untangling one you inherited, or putting a model into production? Write to me and we will find a slot.