MACER PARK

The Problem

Global-e operates a cross-border e-commerce platform. Technical support requires investigating across 10+ disconnected systems: internal tools, payment providers, databases, observability, ticketing and documentation, email logs, and merchant stores. Investigations relied on tribal knowledge that was hard to transfer.

The Solution

I built one MCP server that brings the whole investigation lifecycle behind a single interface. It began as a multi-service stack in Python and Go. I rebuilt it as a single TypeScript MCP server, which cut the number of moving parts and widened what it could do.

Key Features

Unified Order Investigation

A single composite tool pulls the complete picture: pricing, payment transaction history, delivery timeline, refund chain, and merchant configuration, with automatic discrepancy detection. Cross-system investigation went from 2+ hours to under 60 seconds.

Domains Covered

On any given day the cases touch pricing, payments, duties and taxes, shipping, customs documents, returns and exchanges, refunds, fraud and risk reviews, merchant settings, and the replies sent back to merchants. Each lives in a different system with its own API and data format. The tool brings them into one view. When a merchant reports an issue, I count it across every merchant before escalating, so the escalation carries the real scale.

From Mistake to Rule to Hook

When a mistake recurs, the system codifies it: write a feedback memory → add a YAML gate → install a pre-tool-use hook → block the failure mode at the tool boundary. Before any customer-facing response goes out, a checker validates it against those learnings: prohibited terminology, escalation-path correctness, factual accuracy, and tone per recipient.

How Customer Data Is Handled

Investigation queries run on read-only database connections. Customer name and address fields stay masked and are decrypted one record at a time, only on explicit request. Small extraction tasks on decrypted text run on a local model with no network egress, and masking uses deterministic regex.

Rebuilt from a Bare Machine

In mid-2026 a hardware failure during a motherboard swap wiped the machine. I rebuilt the whole system from a bare repository in days. Backups came first: an encrypted archive plus a recovery rehearsal that proves the backup actually restores. I used the reset to redesign the routing layer, data-format conventions, and path handling. A golden-master harness now captures each tool's output byte-for-byte, so a refactor that changes behavior by even one byte fails the gate.

What I Learned

The earlier 7-service architecture was technically more impressive. The later one is the one I can still change safely, which turned out to matter more.

Giving an agent tools is half the work. The other half is codifying how it should behave: what to verify before concluding, and when to ask instead of guessing.