AI & agent-based systems
LLM agents become useful when they can act on real systems. I design the architecture and backend that let them do it safely: tools, permissions, approvals, audit trails and operations that survive failure.
What I build
The model is one component. What makes an agent trustworthy is everything around it: well-defined tools, clear ownership of data, permission checks, human approval where it matters, idempotent operations, and a record of what was done and why. That is software architecture and backend engineering.
Which decisions the agent makes, which tools it may call, how context flows, and where a human approves.
Narrow, well-typed tools over your existing systems, exposed through function calling or the Model Context Protocol.
Permissions per tool, approval gates for irreversible actions, and an audit trail of every call and outcome.
Long-running and retryable operations with status tracking, so a dropped connection or a failed step does not leave the system in an unknown state.
Test scenarios for agent behavior, tracing of tool calls, and metrics that show whether the agent is actually helping.
Agents connected to ASP.NET Core services, SQL Server or PostgreSQL data, and background processing you already run.
Experience
Agents rely on the same foundations as these platforms: a control plane, desired state, background processing and safe access to production data.
Independently architected and delivered a plugin management and licensing platform, connecting backend services, administrative interfaces, and customer-side runtime components.
Read the case study 2017 – 2024Contributed to the development and modernization of an enterprise municipal software platform containing more than ten interconnected subsystems.
Read the case studyFrom my work before founding ArtaPattern. Client names and confidential details are left out, and diagrams are conceptual.
How to start
Every engagement starts small and ends with written outputs. Not foundation-model training or machine-learning research.
For a new product or major capability, before serious development starts: clear system boundaries and an implementation direction.
Not included Full implementation, UI design and infrastructure provisioning.
Ask about system designFor a team with a running system, often on .NET, that is unsure about its structure or where to take it next: what to improve, what to keep, and where to start.
Not included Executing the full migration or refactoring, which can follow as a separate engagement.
Ask about a reviewFor slow responses, recurring errors or failing background processing, investigated before committing to a major rewrite.
Not included Ongoing production support or 24/7 on-call.
Ask about an assessmentScope, duration and pricing are agreed after a first conversation, once the problem is clear.
Engineering notes
Probabilistic, deterministic and hybrid execution paths, and a governance plane that decides approvals independently of model reasoning.
Harness engineering: explicit contracts, isolated environments, closed feedback loops and verification gates around coding agents.
Specify intent, establish domain language and build verification in before delegating work to a coding agent.
FAQ
Well-defined tools over your systems, permission checks per tool, human approval for irreversible actions, idempotent and retryable operations, an audit trail, and evaluation. Those parts decide whether an agent is safe to use in production.
Yes. The usual approach is to expose narrow tools over your existing services, through function calling or an MCP server, instead of giving the agent direct database access.
Yes. I design and build Model Context Protocol servers and tool APIs that let agents use your systems with explicit permissions and logging.
With one meaningful scenario: a single task the agent completes end to end, with its tools, guardrails and failure handling in place. A System Design engagement covers that first scenario.
Yes. I use AI coding agents across discovery, planning, test-driven implementation, review and regression verification, with an engineer reviewing changes before they merge.
Start a conversation
Describe the system and the problem. I'll reply by email with questions or a suggested starting engagement.