Artificial intelligence creates value when it becomes part of a reliable software architecture. A capable language model on its own does not solve a business problem — it has to work with domain logic, data, existing systems and clearly defined processes.
EicDevCon develops AI-enabled applications from the initial business requirement to an operational software system. Traditional algorithms, machine learning, natural language processing and large language models can be combined so that each task uses the technology that actually fits it.
From business requirement to AI application
Every project starts with the process, not the model. Which tasks should be supported or automated? Which information has to be identified? Where does interpretation genuinely help, and where does deterministic processing give the more dependable result?
The architecture follows from those answers. It can combine text recognition, document classification, information extraction, rule-based processing, machine learning and locally or externally operated language models.
The complete development path can be covered from a single source:
- analysis of the business process and its requirements
- software and AI architecture
- preparation of structured and unstructured data
- text recognition and document processing
- language processing and entity recognition
- development and training of classification models
- integration of large language models
- backend, interfaces and user interfaces
- integration with existing applications and business processes
- operation of local, cloud-based or hybrid AI infrastructure
Hybrid AI instead of "an LLM for everything"
Not every task requires a large language model. Clearly defined information is often processed faster and more predictably by rules or conventional logic. Specialised classifiers identify document types, while language processing extracts entities and domain-specific detail.
Language models are added where semantic understanding, wider context or flexible interpretation provides a measurable advantage.
The result is a hybrid system in which rules, established models and generative AI each do what they are suited to.
Local AI, cloud AI or hybrid
AI applications do not require a single deployment model. Depending on data protection, infrastructure, cost and performance requirements, models can run locally, through external services, or across both.
Local language models can be operated inside a controlled infrastructure using Ollama or LM Studio. External interfaces can be integrated where their capabilities add something. Combining the two is equally possible: sensitive processing stays inside the organisation while other workloads use external models.
Note
Deployment models
Local – models in your own infrastructure, for example via Ollama or LM Studio. Cloud – external interfaces, where their capabilities add value. Hybrid – workloads distributed according to data and requirement.
For confidential documents in particular, local processing opens up an additional architectural option.
Document intelligence
One focus area is the automated processing of documents and other unstructured information. Such systems can ingest PDF files and scans, extract content through text recognition, classify document types, identify organisations and reference numbers, analyse content semantically, produce structured data, route documents according to business rules and hand results to downstream systems through interfaces.
This makes it possible to automate document-centric workflows step by step where they previously required manual work throughout.
AI as part of professional software engineering
An AI component does not replace conventional software engineering. Authentication, authorisation, interfaces, data storage, observability, error handling, user interfaces and integration remain essential parts of a production application.
AI engineering therefore sits alongside software engineering, system integration and infrastructure operations. The goal is not an isolated demonstration but software that fits into real processes.
Reference project
For a debt counselling workflow, a custom business application was developed for the automated analysis and classification of incoming creditor documents. It combines text recognition, rule-based processing, language processing, machine learning and language models in a single pipeline; alongside external services, models can be operated locally through Ollama or LM Studio.
The case study describes the architecture and pipeline in detail: AI-Enabled Document Analysis and Process Automation.