Multi-Agent AI App for Personalized E-Commerce Experience

Before that vision could be realized, however, the investor needed a proof of concept that the platform was feasible. They were recommended that the best place to start would be to develop e-commerce features. The goal was to create an automated buying experience powered by AI. Users would use a chat to describe what they wanted and receive personalized recommendations of furniture and lighting fixtures, place orders, and expect deliveries without interacting with the vendor. The experience was designed to feel closer to working with a personal interior designer than shopping through an online store.

LLM-based chat couldn’t deliver the experience that the platform aspired to provide

The first version of the platform looked promising on the surface. It featured a polished frontend and an AI-powered chat interface. But underneath, the system relied on LLM-backed conversations embedded into a traditional architecture that sent user queries directly to language models without modification or proper context management.

As a result, the chat produced generic responses that failed to solve real user problems or deliver the level of personalization required by the investor. Requests such as redesigning a room, matching products to personal style and availability, or completing a purchase required more than a single AI response.

Conversations also lacked memory and structure. The system couldn’t reliably maintain context, adapt to users, or stay within a well-defined role. The chat felt like an automation layer on top of a standard online store. It was not a compelling alternative to online shopping through a standard UI.

DAC.digital was asked to step in and plan the development of the platform that would meet the investor’s criteria. The collaboration began with an AI Workshop that defined the plan to rebuild the platform around agentic AI principles.

DAC.digital rebuilt the platform around an innovative agentic AI architecture that replicated how real teams work

Drawing on experience with distributed systems, our developers designed a multi-agent architecture modeled on how effective teams operate in real organizations.

Just like in a company, each agent was assigned a clear role and specialization, such as planning, delegation, execution, or reporting. Agents were organized into agent teams that were coordinated by a concierge agent. Like a frontdesk of any organization, this agent received user queries and could resolve simple tasks. Complex requests, however, like those requiring multi-step reasoning, many skills or interaction with multiple systems, were delegated further into the agent ecosystem and were passed to a planner agent, which broke the user’s goal into a step-by-step execution plan.

Specialized agents, like trained experts in an organization, were assigned tasks they were a great match for, such as pick products from the catalog or generate an image. They could collaborate with each other via A2A protocol and if they needed access to external systems or non-agentic logic, they used MCP. Once they completed the task, they sent the results to a reporter agent responsible for generating a response that was sent to the user.

For example, a user uploaded a photo of their living room and asked for a chair recommendation. The concierge received the request and determined that it would require multi-step reasoning and found a team that would be a great fit for this task. The request was then passed to a planner agent, which created a plan that involved image understanding, product matching, and image manipulation. It distributed the tasks among specialized agents to propose products from the real catalog, tailored to the user’s preferences and space. The results were then synthesized by a reporter agent and returned to the user through the chat interface. The user could refine their request or proceed with a purchase.

Agents could be added or removed at runtime, and if one becomes unavailable, another can take over

What was innovative about this approach was that the AI system was designed as a living network of agents who can join the agentic structure and execute tasks without causing service disruptions. To participate, agents simply registered themselves in a central agent registry backed by Redis. During registration, each agent presented an agent card, a structured profile describing its skills and availability. Essentially it was like “clocking in for work” by a human employee.

The result was a microservice-inspired ecosystem that operated continuously without downtime even if the investor wished to add more agents, introduce new features, and respond to evolving user demands.

The chat could finally give personalized recommendations

Thanks to the new architecture, the platform could deliver recommendations that felt truly personal. The system had access to user profile and retrieved context in real time via RAG (retrieval-augmented generation), so it could consider the user’s previous interactions, product availability, and preferences when crafting suggestions. The AI team also used prompt engineering to make sure that the chat responses were accurate, context-aware, and human-like.

The solution leverages vendor-agnostic frameworks

By avoiding lock-in to any single platform or provider, the system could integrate seamlessly with multiple vendors, marketplaces, and services. It also allows for the platform to evolve as technology changes which is especially important in engineering AI solutions where technology evolves all of the time.

Agentic frameworks:

• Agno – generic agent framework
• Temporal.IO – orchestration of complex, non-deterministic tasks
• Milvus – vector database
• LlamaIndex – document parsing
• OpenAI models  – reasoning engines
• Google Gemini’s Nano Banana PRO – image analysis tasks

Observability:

• Phoenix – monitors agent operations, delegations, and task completion

Testing: 

• Prompteval – for testing conversational integrity, checking if that agent remains on-topic, and does not hallucinate

The foundation that DAC.digital built now positions the investor to realize the vision of a city-wide operating system

The reengineered platform now provides a seamless, responsive experience where users always have an agent to interact with, while complex tasks are executed by other agents. The agentic AI workflows are flexible, easily extendable, and capable of evolving alongside the business. And that is happening right now, because the platform is about to scale beyond the e-commerce focus. It can incorporate additional user groups, new features, and support seller activities that the platform planned to enable like product promotion and ad generation.

Ready to start a project with us?