What I shipped
- Architected a high-throughput ClickHouse & dbt ELT platform across 100+ tenants, replacing legacy Python ETLs — nightly runtimes 5 hours → 15 minutes (20× speedup) and analytical query latency from 15+ seconds down to 1–2 seconds.
- Decommissioned heavy db.m7g.8xlarge RDS OLAP instances by provisioning ClickHouse on AWS ECS via AWS CDK with EC2 capacity providers — database costs down 85% (~$25,000/year) and storage down 75% (1 TB → ~200 GB) via columnar compression.
- Architected a zero-downtime Postgres → ClickHouse CDC pipeline (PeerDB on ECS Fargate) across 4 microservices, replacing CPU autoscaling with step-scaling on custom CloudWatch metrics to keep replication-slot lag under 1 GB across multi-tenant mirrors.
- Engineered an idempotent upsert migration framework with a custom Python query-translation engine that rewrites legacy SQL to ClickHouse SQL — seamless cutovers for 100+ Superset client dashboards.
- Developed custom Model Context Protocol (MCP) servers for Apache Superset and ClickHouse, accelerating agentic AI workflows for automated SQL translation, live query execution, and system debugging.
Stack
ClickHouse 
dbt 
Python 
PostgreSQL PeerDB AWS 
Superset MCP ClickHouse dbt Python PostgreSQL PeerDB AWS Superset MCP