Architecture and the build

Maciej Kaczor

Chief Technology Officer, Ovidius AI

Maciej owns the architecture and the software on every Ovidius build. Builds run inside your environment on self-hosted n8n, connect to your CRM, ERP and databases, and are tested against a staging replica before they touch live records.

Read Maciej's articles Maciej on LinkedIn

Maciej Kaczor, Chief Technology Officer of Ovidius AI

What Maciej writes about

Maciej writes about what sits between a model and your live systems: context layers and the business rules inside them, MCP servers and how to run them safely, how agents connect to ERP and CRM data, chatbots that act on records instead of answering FAQs, and the guardrails that stop an agent before it acts on a bad answer. He also publishes the benchmark Ovidius ran on which Claude model to hand delegated build work to.

The work behind them

Enterprise AI chatbot AI invoice automation Context layer Forge AI Proof of concept

Background: Performance and process optimization, from refining search algorithms to architecting engineering workflows, before leading engineering on agentic workflows and LLM-powered systems.
Experience: 4 years in AI

Articles by Maciej

Five guardrail checkpoints outside the model: input, tool result and output catch bad text, while the highlighted pre-tool and environment checkpoints catch bad actions before a system changes.

AI Guardrails: How to Stop an Agent Before It Acts on a Bad Answer

Where AI guardrails sit in an agent, why the check before a tool call matters most, what to validate there, what happens when a check fires and how to test it before launch.

Scorecards for Haiku 5.5, Sonnet 5.5 and Opus 5.5: 14, 4 and 0 of 18 agents, with quality, perfect scores, median time and an effort heatmap for each

Claude Code Subagents: Haiku vs Sonnet vs Opus, Benchmarked

We graded Haiku 5.5, Sonnet 5.5 and Opus 5.5 on 23 kinds of delegated Claude Code work. 14 of 18 agent definitions moved to Haiku, and none needed Opus.

One request built two ways: 100k tokens with everything in the prompt, against 4.1k tokens of engineered context, highlighted.

What Is Context Engineering, and Why Enterprise Agents Fail Without It

What context engineering is, how it differs from prompt engineering, the four ways an agent context fails on enterprise data, and the techniques that fix each one in production.

A four-layer stack from AI agent down to the warehouse, with the context layer highlighted above the semantic layer.

Context Layer vs Semantic Layer: What AI Agents Need From Your Data

A semantic layer fixes what your numbers mean; a context layer decides what an agent may read and assembles it. How each works, how each fails, and when an agent needs both.

Five steps from one workflow to an agent program, prompt and response, task automation, tool-using agent, multi-agent workflow and agent program, with the agent program highlighted.

The Agentic AI Roadmap: From One Workflow to a Managed Agent Program

How to grow one production agent into a managed program: the five steps, the architecture a multi-agent system needs, governance matched to what each agent may do, and how to run it.

What a customer service AI agent resolves alone, highlighted, order status, returns inside the window and address changes, set against what it hands off to your team.

Customer Service AI Agents: What Can They Actually Resolve Without a Human?

Which customer requests an AI agent can close alone, which belong with a person, the four checks behind every handoff, and what one production chat agent resolved.

Four levels of enterprise chatbot use case, answer, look up, act and orchestrate, with level 3, act, highlighted as the step where the chatbot changes a record.

Enterprise Chatbot Use Cases That Go Beyond FAQ Answers

Four levels of enterprise chatbot use case, from cited answers to multi-step workflows, with examples by department, what each needs from your systems, and where to start.

SharePoint, Google Drive and Notion sources feed an orange answer step that links the exact section, delivered in Slack and Microsoft Teams.

How to Build a Knowledge Base Chatbot Your Employees Will Actually Use

The parts of a knowledge base chatbot, which documents to load first, how to keep permissions intact, when it should say it does not know, and how to test it before launch.

Requests from Teams and Slack pass through an orange integration layer that acts as the user, applies rules and approvals and logs every step, before reaching SAP, Odoo and Salesforce.

Enterprise AI Architecture: How Agents Connect to Your ERP, CRM, and Core Data

Five layers between an agent and your ERP or CRM, a read, draft, write permission model, rules kept out of the model, and the log fields CISA's Logging Reference Architecture asks for.

Three figures from one invoice agent run: a 1% tolerance rule, an invoice 2.2% over its purchase order routed to the AP lead, highlighted, and five checkpoints on every run.

Enterprise AI Agents in Production: What Works Inside Real Systems

One €18,400 invoice against an €18,000 PO, traced through five checkpoints from extraction to the run log, plus the six monitoring categories NIST AI 800-4 sets for live agents.

Claude, ChatGPT and n8n agents reach ERP, CRM and docs servers only through an MCP gateway that checks permissions, asks approval on writes and logs every call

MCP Security: How to Run MCP Servers Safely in the Enterprise

Six public incidents show where MCP deployments break. The controls that stop each one, how MCP authorization works with OAuth 2.1, and what an MCP gateway adds.

Claude, ChatGPT and n8n connect through one MCP layer to ERP, CRM and file systems: 9 pieces instead of 18 connectors

What Is an MCP Server? The Model Context Protocol, Explained

An MCP server wraps one system, such as an ERP or a file store, so any AI app that speaks MCP can use it. How hosts, clients and servers fit together, and how MCP differs from an API.

Five context layer tiers built in order, ingestion, knowledge graph, sandboxing, delivery over MCP and permissions, with permissions checked before retrieval highlighted.

How to Build an Enterprise Context Layer: The Five-Layer Blueprint

The five layers of an enterprise context layer, what each one is built from, how to settle a term like revenue, which layer fixes which agent failure, and the order to build them in.

EHR, clearinghouse and billing system data pass through an orange context layer that applies a field list per task, masks identifiers and logs every retrieval, before a model sees anything.

Architecting an Enterprise Context Layer for Protected Healthcare Environments

What HIPAA asks of an agent that reads patient data, how a context layer limits each task to the fields it needs, where PHI leaks in an agent pipeline, and a checklist to start with.

A SAP Joule Studio agent calls n8n over MCP, and n8n runs the steps for a supplier follow-up and unmatched invoice triage.

n8n + SAP Joule Studio: Integration Guide

How a Joule Studio agent calls n8n over MCP through a BTP destination, which of n8n's three MCP options to use, the setup steps, and which work belongs in Joule, n8n or SAP.

Business rules pasted into a system prompt, weighed loosely with no trace, set against rules stored as versioned records in a context layer, highlighted, where an engine returns the verdict and each run is logged.

Teaching AI How Your Business Works: Mapping Rules into the Context Layer

Policy pasted into a prompt gets less reliable as it grows, as Chroma's study of 18 models shows. Store each rule as a versioned record with an owner, and let a rules engine make the call.

More from the Ovidius partners

Each partner writes about the stage of the work they own on client builds. Meet the team

Owen BoesveldChief Executive OfficerArticles and background

Jason ReynoldsChief Operating OfficerArticles and background

Put your first workflow on the board

Answer a few questions about your team and what you want automated, then Jason scopes it with you on the call.

Owen, Maciej, Jason, Oskar and Ben