An integration architecture, not another isolated AI tool
We research how existing applications, AI systems and business workflows can work together through a reusable architecture, rather than a separate custom connection for every new use case.
Ongoing internal research · Since 2021
AEGIS
AI-Enabled Generic Integration System
Connecting the systems businesses already depend on with the possibilities of AI — without losing control of how the work gets done.
We research, develop and test our own systems, continuously revisiting the architecture as AI technologies evolve.
AEGIS is our internal research activity. It is not a paid service or a product offered for sale.
The research question
How can AI join a business process without taking it over?
Project AEGIS researches a generic architecture that enables existing software systems to participate in AI-enabled business processes through semantically enriched interface definitions, requiring only minimal modifications to existing applications and preserving deterministic process execution.
In everyday terms: describe what your systems can do and what their operations mean, connect them to AI, and keep the business rules, permissions and responsibility clear. This is a research objective — not a claim that every integration challenge has already been solved.
Interfaces that explain meaning
An interface should describe more than how to call a system. We explore definitions that also explain what an operation means, what information it needs, its constraints and what its result represents — so AI can work with business systems more appropriately.
Keep existing applications in place
The aim is to expose the capabilities of existing software with minimal modifications. Businesses should be able to explore AI integration without rebuilding every application or discarding the systems they already rely on.
Predictable processes around variable AI
AI responses can vary. We research deterministic orchestration: the surrounding workflow follows explicit rules, checks and approved steps, even when an AI component does not produce the same answer each time.
Build frameworks and working systems
We develop the architecture and core Java framework ourselves, implement experimental systems and test them. The work draws on Java, Python, C++, Spring Boot, REST and LLM APIs, enterprise integration and distributed systems.
Validate, correct and keep learning
Runtime validation checks results as a process runs. Adaptive prompting and automated error-correction mechanisms are investigated alongside limits and failure handling — testing where these techniques help and where human review is essential.
Long-term research. Constantly renewed.
The 2010s · Foundations
Our exploratory work began with neural networks and emerging language-model research, alongside a long-standing interest in enterprise integration — before the recent wave of widely adopted generative AI tools.
Early 2021 · Continuous research
AEGIS became a continuous company research activity, focused on a generic approach to integrating AI into existing enterprise systems and business processes.
Today · Build, test, refine
We develop working systems, test integration approaches and revisit our assumptions as models, protocols and tools change. Research stays active, rather than stopping at a demonstration.
Ways AI and software work together
More than a conversation with a model
Our research follows a broad and evolving range of integration approaches. These are complementary building blocks — some are protocols, others are architectural patterns or ways of organising work. There is no single connection method that fits every business.
Model APIs & REST services
Software sends a defined request to an AI model and receives a response. REST is a common web-based way to make these calls and connect existing business applications.
Tool & function calling
AI proposes a specific action, such as looking up an order. The application checks the request and permissions, performs the action and returns the result. AI does not receive unrestricted access.
AI agents
An agent combines a model with tools and a workflow to pursue a task over several steps. We examine how to keep those steps bounded, observable and subject to approval where needed.
MCP — Model Context Protocol
A shared protocol for connecting AI applications to tools, data and reusable prompts. It can reduce the need for a different connector for every application, while access rules still need to be enforced.
A2A — Agent-to-Agent Protocol
A protocol for agents from different systems to exchange information about their capabilities, delegate tasks and track progress. Cooperation does not remove the need for ownership or checks.
Agent skills
Reusable packages of instructions, supporting resources and sometimes scripts that guide an agent through a particular kind of work. They provide task-specific know-how, rather than a communication channel by themselves.
RAG — Retrieval-Augmented Generation
Relevant information is retrieved from an approved knowledge source before AI responds. This can ground an answer in company documents, but does not guarantee that every answer is correct.
Events, messaging & webhooks
A business event — an incoming document or a changed order, for example — can trigger an AI-supported step. Queues and notifications let systems cooperate without waiting on one long conversation.
Multimodal & real-time interfaces
AI can work with text, images, documents, audio and, where supported, video. Streaming and real-time connections allow results or voice interactions to arrive progressively rather than only at the end.
Structured outputs & workflow orchestration
AI returns information in an agreed format that software can validate. A workflow then decides what may happen next, including approval, retry or escalation. A well-shaped answer is still not proof of a correct answer.
Across providers, not tied to one
We test major AI providers and model ecosystems as part of the research, looking beyond a convincing answer to how they behave inside real software and business workflows.
The landscape includes OpenAI, Anthropic (Claude), Google (Gemini), Microsoft (Azure AI), Amazon Web Services (Amazon Bedrock), Meta (Llama), Mistral AI, Cohere and DeepSeek. Some provide models directly; others provide platforms for accessing multiple model families. Our coverage evolves as technologies and access options change.
Integration behaviour
Tool use, structured responses, context handling and cooperation with existing applications.
Control and reliability
Validation, failure handling, repeatability of the surrounding workflow and human approval.
Business suitability
Quality, security, confidentiality, response time, operating cost and deployment constraints.
Provider names identify the technology landscape, not partnerships or endorsements. Capabilities depend on the model, version and deployment.
Research that strengthens our client work
AEGIS gives us a practical foundation for helping businesses with AI-related tasks, projects and decisions: understanding the options, connecting existing systems, inspecting a proposed solution or developing and testing an implementation.
That experience informs both complete project delivery and the work of our senior professionals in client teams. Any client assignment has its own agreed scope; AEGIS itself remains our ongoing internal research.
How we work with AI
AI improves our work. Humans own it.
Whichever way you work with us — a senior professional placed in your project, or a complete project taken on by us — our professionals use AI and AI-driven tools without exception, to improve the quality, efficiency and completeness of their work. They are all up to date with the latest developments in this area.
But they never hand over their work to AI. They implement it themselves — and they take responsibility for it.
AI is one of our areas — not our main area. IT system development and integration are our thing, with or without AI. And within AI itself, our main subject is not developing an AI service or a model of our own: many highly capable researchers and companies around the world are already doing that, and we appreciate their work. Our main subject is what and how you and we can do things with these AI services — in practice, and realistically.
"Those who cannot do something without AI cannot do it with AI either. AI can only improve, but not replace."
In all our business, the human is the main component. Not because AI capability is untrustworthy — but because human capability is irreplaceable.