Artificial intelligence. Engineered with purpose.
We turn AI into a productive part of your software. With deep engineering expertise, our own powerful servers and an architecture that keeps you in control.
ETZENSPERGER INFORMATIK AG
Illustration of a vector database with connected topic clusters and keywords. The original etz.ch wordmark floats at the centre as a three-dimensional outline in its original colours. Each topic cluster connects to the cyan dot in the etz.ch logo. No actual customer data is shown. Software: APIs, TypeScript, SaaS. Knowledge: RAG, Embeddings, Retrieval. Security: Code Review, SAST, CI/CD. Infrastructure: GPU, Inference, On-Premises. Models: Qwen 3.8, GPT-6 Astra, Claude Opus 5. Data: Documents, Context, Metadata.
Etzensperger Informatik AG combines extensive software project experience with practical AI expertise. We connect language models, data and business logic where they create measurable value.
Retrieval-augmented generation (RAG) connects language models to your approved documents and data. Embeddings and vector search supply relevant context, sources make answers verifiable, and access rights apply at retrieval.
RAG / EMBEDDINGS / RETRIEVALAnalyse documents, structure information and prepare processes. We integrate AI into applications and APIs, with structured outputs, validated tool calls and clearly limited permissions.
API / TOOL CALLING / WORKFLOWSA prototype is the beginning. Representative test cases, versioned prompts and targeted evaluations help assess answer quality, latency and cost, making model changes and new features traceable.
EVALS / OBSERVABILITY / LIFECYCLEPrivate AI / Qwen 3.8
We run Qwen 3.8 on our own powerful servers. With this local deployment, sensitive data and source code are processed within our infrastructure without being sent to external model providers for inference. AI performance comes with a clearly defined data flow.
Our deployment pipelines use our locally hosted Qwen model to inspect code changes for potential vulnerabilities. It assists with contextual review of data flows, input validation and authorisation logic, bringing suspicious code to light before release.
Select a step to explore the review workflow.
Commit & diff
Qwen · local
Engineering team
Tests & deployment
The process starts with a code change. The diff and relevant application context show which logic changed and how data flows through the affected functions.
A user search inserts input directly into an SQL query.
The changed function and its data flow are assembled for local analysis.
Our locally hosted Qwen model examines the change in context. It flags suspicious data flows and supports engineering review with a location and an explanation.
Unchecked input could alter the SQL query structure: a potential SQL injection.
The finding goes to the engineering team. A model response alone does not decide whether to release.
Engineers validate the finding in its application context. In this example, a parameterised query replaces direct concatenation, and a regression test checks the correction.
− query("SELECT * FROM users
WHERE id = " + input)+ query("SELECT * FROM users
WHERE id = ?", [input])After the fix come further reviews, automated tests and the required security checks. The change proceeds to deployment only when the defined release criteria are met.
Special characters are passed as data. The input cannot alter the structure of the parameterised query.
Review complete, tests passed and required checks satisfied: the corrected change can be deployed.
Engineers validate AI findings. They complement code reviews, automated tests and security scanners. An AI review without findings does not guarantee vulnerability-free code.
We have also delivered software projects using OpenAI and Anthropic’s Claude. You determine the permitted integration approach and data processing location. We design the architecture around those requirements.
Qwen on our own servers, for sensitive data, internal code and applications with clearly defined data flows.
Qwen 3.8GPT-6 Astra from OpenAI or Claude Opus 5 from Anthropic in your application. We align the model, API, data categories and contractual terms with your use case.
GPT-6 Astra / Claude Opus 5Local processing for confidential content, external models for explicitly approved tasks. Routing follows agreed data rules.
Qwen 3.8 + GPT-6 Astra / Claude Opus 5Before implementation, we check the required location against available provider regions and contracts, including storage, logs, backups and subprocessors. Not every model and region combination is available.
Personal data and secrets should enter context only when the task requires them and processing is approved. Data minimisation starts before the prompt.
A language model is not access control. Enforce roles, tenant isolation and permissions in the application and during retrieval.
Documents and websites can contain malicious instructions. Limit tools, validate calls and require approval for critical actions to reduce prompt injection risks.
Models can generate plausible but incorrect statements. Use sources, tests and human review before applying results to code or business processes.
Prompts, embeddings, answers and logs can contain confidential information. Access, encryption, retention and deletion belong in the operating plan.
Local hosting alone does not make an application secure. Updates, monitored interfaces and renewed evaluations after model changes remain essential.
From a good idea to productive AI
Bring your use case. Together we will clarify value, data requirements and integration, then define a concrete next step.