ApparelMagic AI Agents

I architected and led the build of ApparelMagic's AI agent platform, working collaboratively with the team and incorporating hardening and final engineering passes from ApparelMagic's CEO/lead engineer. AI Designer is covered in its own case study; this one is about the agents.

The core problem was turning a system of record into something that could monitor operations and act on them. I built configurable agents with their own instructions and business context, then wired them to pre-built tools that read ApparelMagic data and return real operational insights: stock and backorders, customer and account risk, invoices and margins, product demand and assortment.

A lot of the hard learning was in the execution model. Agent jobs needed queue-based async execution so scheduled and long-running work could happen reliably without blocking the app. That meant designing for retries, concurrency, and progress that operators could actually trust when agents ran overnight or across large datasets.

Tool design mattered as much as the models. Rather than dumping raw ERP tables into prompts, agents called purpose-built tools against inventory, sales, finance, and product-development data, then reasoned over structured results. That kept answers grounded and made it possible to draft actions the business could review.

One of the more useful patterns was LLM digestion of those results into customer-facing work: data-aware bulk email drafts for sales and customer engagement, plus custom reporting that turned agent findings into something a human could send or act on. Permissions, approval gates, and auto-approval controls sat in front of anything that wrote back to the system.

Shipping this taught me how much agent quality depends on tool boundaries, job orchestration, and human-in-the-loop controls, not just prompt quality.