Work / Business analytics

From a reporting bottleneck to business answers.

Reporting in seconds, backed by an analytics foundation and the official Superset MCP service.

Seconds

instead of hours or days

Problem

At a US consumer app, a business question often turned into an engineering task: filter data, prepare an export and plot the result. Reports took hours or days, slowing decisions and interrupting engineering work.

Constraints

The work needed to support nontechnical team leads and connect to live business data. An easier interface would only be useful if the underlying reporting system could supply the answers.

What I did

I built a near-real-time analytics stack with ClickHouse and Apache Superset, including KPI dashboards and scheduled email reports. The work included internal infrastructure and a Superset deployment on AWS Fargate. I chose ClickHouse for a workload of millions of events a day. The queries themselves took 5 to 7 minutes on a MySQL box with 80 to 90 GB of RAM; they now return in seconds. That is query time, separate from the hours or days a report used to take. I dual-wrote into the new store rather than using change data capture, because it was simpler, and I used a JSON column so event shapes could vary. Superset is there for drill-downs.

I connected the official Superset MCP service, which is part of the Apache Superset codebase, so nontechnical team leads could ask questions of live business data in plain English. I did not write a custom MCP build. The dashboards and the language interface shared the same purpose: make business data accessible without an engineer preparing each answer.

Result

Routine reporting moved from hours or days to seconds. Team leads could answer questions themselves instead of waiting for a manual export. The non-technical moderation team now asks its own questions, using reusable chats and dashboards.

Tradeoffs

Natural-language access adds another interpretation layer. Clear metric definitions, query complexity and source freshness still determine whether an answer is useful. The analytics foundation remains essential even when the interface becomes conversational. When the model is not sure, it says so, and the answers need human review. The path runs in the same VPC. The data is PII-free, with only user IDs.

Stack

ClickHouseApache SupersetMySQLMCPAWS Fargate

Explore more problems I've solved

Bhavin Kamani

A small command line for a curious visitor.