Descriptions:
Sumeet Marwaha, a data leader with experience at Brex, joins Peter Yang to build a working AI data analyst using Claude Code in a live session, mapping out how AI can augment the complete analyst workflow: monitoring dashboards, investigating anomalies, exploring transaction-level data, constructing business narratives, and estimating impact. The session uses Claude Code in bypass-permissions mode integrated with Brex’s internal MCP (Model Context Protocol) server for structured data access.
A central technical challenge Marwaha addresses is context window management for SQL-heavy analysis. When Claude runs an unbounded query returning millions of rows, it can exhaust the available context or — more insidiously — mistake a limit-1000 sample for a complete dataset in subsequent reasoning steps. His solution is explicit prompt-level instructions directing Claude to track token usage across a chain of analysis steps and to remember that prior queries were sampled. He also identifies semantic scoping as critical: providing Claude only the single customer segmentation relevant to a given analysis, rather than all eight available dimensions in Brex’s core data table, prevents the model from inconsistently selecting different segmentations across runs.
Marwaha discusses how Cursor has emerged as the dominant coding tool among Brex’s startup customers and enterprise clients alike. He contrasts Claude’s ability to simultaneously cross-reference Slack threads, Linear tickets, and raw transaction data — mirroring how experienced data scientists actually contextualize findings — with earlier AI tools that could only assist with SQL debugging. The session positions Claude Code with MCP as the first tool he believes could realistically deliver end-to-end analyst automation, from anomaly detection through impact sizing.
📺 Source: Peter Yang · Published January 18, 2026
🏷️ Format: Hands On Build
