Project
Merchmind Multi-Agent Commerce System
Four domain agents and an orchestrator coordinate assortment, pricing, inventory, and category decisions with replayable conflict resolution.
Multi-Agent · Orchestration · LLM/Tools Layering · ERP Concepts · Explainability
Product Snapshot
- Role
- System design and demo implementation, team project
- Users
- Commerce operators coordinating assortment, pricing, inventory, and category strategy.
- Stage
- Runnable demo, system architecture, and presentation chain completed in 2026.03-2026.06.
- Focus
- Multi-agent orchestration, deterministic tools versus LLM reasoning, conflict arbitration, and explainability.
- Validation
- Team records report simulated conflict rates falling from about 30% to below 5%, with the trial cycle shortened from 14 to 7 days.
- Public Proof
- Runnable demo, six arbitration rules, replay traces, and complete presentation materials; reported metrics were not independently rerun for this portfolio.
Next Step
The problem
Commerce decisions rarely belong to one function. Assortment, pricing, inventory, and category strategy affect one another, so independently optimizing agents quickly create conflicts that are difficult to explain.
System design
We separated the system into Category, Pricing, Selection, and Inventory agents, with an orchestrator coordinating context, conflict detection, and execution order. Deterministic calculations stayed in tools, while the LLM handled open-ended reasoning.
Six arbitration rules resolved competing margin, stock, price, and risk objectives. Each run preserved inputs, disagreements, arbitration, and final actions for replay.
Result
The team delivered a runnable demo and a complete presentation flow. Project records report simulated conflict rates declining from roughly 30% to below 5%, and the trial cycle moving from 14 to 7 days. These are team-project results, not independently reproduced production metrics.
What I learned
The value of a multi-agent system comes from boundaries, deterministic tools, conflict protocols, and auditability, not from the number of agents.
Configure the public Giscus environment variables to open discussion on the public site.