Multi-agent AI-support facilitating highly customized order completion.

The challenge

The challenges of making AI a printing expert

It is typical for people to get overwhelmed when options are too abundant. The same applies to artificial intelligence, which may struggle if it has to pull options from an excessive variety of components to choose from. Such are the challenges of applying LLMs to specific business needs.

The AI agent for Mixam had to be engineered to create order specifications that would be validated when taking orders in. With challenges like this, engineering expertise is key to blending the indeterministic nature of language models with the rule-based backbone of the final delivered solution. This goes against the common perception that the deployment of AI agents is straightforward — it almost never is in a business context, and it requires an art of engineering and expert knowledge to make it work as a reliable part of any business solution.

This also applies to securing the solution so that it is not thrown off and begins talking garbage, protecting it from a drop in reliability and from increasing hallucinations. Guardrails forbid the agent from picking up topics that are not directly related to printing orders. As a result, the agent refuses to give restaurant recommendations or offer cupcake recipes. When it acts as a self-publishing advisor, it focuses solely on helping the end-user make the right selection of Mixam options so the final print meets the customer's needs.

How it works

From chit-chat to expert advice

The challenge of creating a true agentic AI printing expert required overcoming typical issues related to large language models, such as hallucinations. Specifically, to prevent the bot from playing the guessing game and instead offer what is actually printable requires the narrowing down of its options to concrete product specifications. Vstorm achieved this by having the AI agent use Mixam product options, pulling them from Mixam's publishing systems with API calls.

This grounded the solution in choosing existing combinations of cover, page, and binder options. However, having a simple AI bot use the components of Mixam products and build an order from them was not enough — the solution needed to know how to recommend the right combinations of those components to potential customers. This was accomplished by supplying the agent with a Mixam knowledge base that combines both best practices and the most common choices that users have made and proven happy with.

AI tooling necessary to cater to a user's request

The Vstorm team built the solution using the PydanticAI Python-centered framework, with FastAPI for inter-application processes and a powerful RAG vector store for matching requests with products based on Mixam's always-up-to-date internal knowledge of their product features.

The choice of this environment was made as it is built with safety as a priority and focuses on strict data validation. As a framework relied upon in the healthcare and finance industries, it was deemed adequate for self-publishing.

Claude Haiku 4.5

Compass, conversational print specification and ordering

Compass guides customers from a plain-language description of a print job to a priced, fully specified product in the basket, including handling modifications to existing orders.

Rendreal's customers specify and price complex print jobs through an automated conversation, removing the need for a designer or sales consultant in the loop while maintaining production-grade accuracy.

Claude product and model used

The solution is built on Claude Haiku 4.5, orchestrated via the Pydantic AI framework. Haiku was selected for its industry-leading quality-to-speed ratio, which is critical for maintaining a seamless, conversational user experience where low latency directly correlates to conversion. The architecture utilizes Haiku as the primary agentic engine, managing a multi-agent system that includes specialized agents for product selection, suggestion generation, confirmation, and conversation titling.

Problem

Specifying a print job is expert work involving interdependent variables — stock, weight, binding, lamination, and format. A single choice made early in the process can double costs or rule out specific finishes later. For non-designer customers, this complexity leads to high basket abandonment or unnecessary reliance on support staff. Rendreal required a solution that could open this professional process to anyone capable of describing their project, arriving at a valid, priced specification without manual intervention.

Solution

Vstorm engineered the Compass agentic system on Claude Haiku 4.5, recognizing that in high-stakes printing environments, reliability and latency are paramount. The architecture is flat and robust: a single agent governs the conversation, with all capabilities — product search, component assembly, pricing, and order management — hanging off it as validated tools.

Claude Haiku 4.5 serves as the cognitive engine for this system. It decides which tools are necessary for each user request, often triggering six or more tool calls in a single turn to assemble a specification. By leveraging Claude Haiku's ability to handle constrained generation, Vstorm successfully bridged the gap between the indeterministic nature of natural language and the rule-based, deterministic requirements of the production printing backend. The system ensures that the agent focuses solely on valid, printable product configurations, preventing the guessing behavior typical of generic chatbots.

Model evaluation

Vstorm evaluated the product selector component on 67 test cases, spanning direct requests, implicit descriptions, and out-of-stock scenarios, totaling 670 runs per model.

Claude Haiku 4.5 matched the highest overall accuracy in the group at 98.51%. It achieved this accuracy in 2.01 seconds per request, outperforming GPT-5.1 (6.31 seconds) and GPT-4.1-mini (2.36 seconds). On strict best-match accuracy, Haiku scored 92.54%, a negligible trade-off for three times the speed of the next-best model.

The evaluation confirmed that Haiku offers the best quality-to-speed ratio for a chat interface, providing the responsiveness required to turn customer inquiries into confirmed, paid orders in real time.

Results

What changed for Mixam's first-time publishers

Mixam's approach exemplifies the industry shift from "AI projects" to "AI products." The infrastructure includes sophisticated monitoring, centralized prompt management, and performance optimization.

This operational maturity — treating AI as production infrastructure requiring monitoring, version control, and staged deployment — distinguishes systems built for long-term business value from experimental prototypes. The repository structure itself tells the story of incremental evolution: test logs, performance benchmarks, shop-specific validation scripts, and careful Git management all point to a team iterating based on data, not assumptions.