Designing AI agents
Exploring principles, patterns and best practices for agents that have clear responsibilities, constrained autonomy, observable behavior and well-defined human oversight.
Agent boundariesTool useMemoryEvaluation
Research & Development · MedTech · Agentic AI
Exploring how Agentic AI can make regulated software engineering leaner, faster and maintainable.
DINISO is an R&D initiative focused on how AI agents, specification-driven engineering and multi-agent workflows can support software teams working in environments where quality, traceability, security and maintainability matter as much as delivery speed.
LLMs•MCP•A2A•AI Agents•Specifications•Regulated Software•MedTech•Maintainable Code•
LLMs•MCP•A2A•AI Agents•Specifications•Regulated Software•MedTech•Maintainable Code•
01 — Research agenda
How can AI accelerate delivery without turning software engineering into a black box?
The goal is not simply to generate more code. It is to investigate how Agentic AI can improve the complete engineering workflow while preserving architectural clarity, reviewability, testability and long-term maintainability.
Vibe engineering
Investigating how natural-language-driven development can evolve from ad-hoc prompting into disciplined engineering: fast exploration, explicit intent, continuous verification and deliberate human review.
Rapid iterationGuardrailsReviewVerification
Spec-driven development
Treating specifications as executable engineering context so that architecture, domain rules, quality constraints, acceptance criteria and regulatory needs can guide both humans and AI agents.
SpecificationsTraceabilityAcceptance criteriaTests
Multi-agent intelligent workflows
Building and evaluating workflows where specialized agents collaborate through LLMs, Model Context Protocol (MCP) and Agent-to-Agent (A2A) communication to perform analysis, implementation, validation and review.
LLMsMCPA2AOrchestration
02 — Engineering hypothesis
AI should amplify engineering discipline, not replace it.
Lean, not careless
Reduce handoffs and repetitive work while keeping architecture, reviews, test evidence and critical decisions explicit.
Agile, not unstructured
Short feedback loops should coexist with specifications, traceability and deliberate quality controls—especially in regulated product development.
Automated, but observable
Agent actions should be inspectable. Inputs, decisions, generated artifacts, validations and handovers should leave an understandable engineering trail.
Fast today, maintainable tomorrow
The success metric is not code generated per minute. It is useful software that remains understandable, testable, adaptable and safe to change.
03 — From experience to research
This R&D direction grows out of years of building real software systems.
My background is in senior full-stack and cloud-native software engineering, primarily with C#/.NET, distributed systems, Azure, web applications and integration-heavy products. In MedTech, that work brought software engineering into direct contact with regulated development, cybersecurity, interoperability, clinical data and quality-oriented delivery.
At Cortrium, I worked on medical-device software and SaaS capabilities around ECG workflows, external integrations and report delivery, while continuously deepening my understanding of standards and practices such as IEC 62304, ISO 13485, ISO 14971, EU MDR and related software lifecycle concerns.
Earlier product and consulting work—including taking ideas through domain discovery, event storming, DDD, prototyping, team formation and production launch—reinforced a recurring lesson: software quality depends as much on how decisions are structured as on how code is written.
The current DINISO research direction is a natural next step: investigating whether AI agents can participate in those engineering workflows without losing the rigor that makes complex and regulated systems trustworthy.
04 — A possible intelligent workflow
From intent to evidence.
One research direction is a coordinated set of specialized agents, each operating with explicit responsibilities and checkpoints rather than one unconstrained code-generating assistant.
Product / domain intent
Capture goals, constraints, risks and user value.
Specification agent
Turn intent into structured requirements and acceptance criteria.
Architecture agent
Propose boundaries, interfaces, trade-offs and implementation guidance.
Implementation agent
Generate or modify code within the agreed constraints.
Verification agents
Run tests, static analysis, security checks and consistency reviews.
Human decision
Review evidence, resolve ambiguity and retain accountable control.
Professional reference
A living R&D portfolio, not just a company landing page.
DINISO is being used to document experiments, architecture ideas, working prototypes and lessons learned at the intersection of Agentic AI and disciplined software engineering. It also serves as a professional reference for roles involving senior software engineering, architecture, AI-assisted development, cloud-native systems and MedTech.
Let’s connect
Interested in software engineering, Agentic AI or regulated product development?
Whether you are hiring, exploring an R&D collaboration, or simply want to exchange ideas about AI-assisted software engineering in regulated environments, I would be happy to hear from you.