AI Engineering — AI that stays in production.
95% of software engineers use AI tools weekly. Between "works in the notebook" and "runs GDPR-compliant for 10,000 users" lies an engineering problem — not a consulting problem. We solve it inside Build Sprints. The following engineering packages are typical sprint content.
Engineering-Pakete
Engineering packages:
AI feature in existing platform
6–12 days. LLM integration into existing SaaS or Mittelstand platform. NestJS or Spring Boot backend, frontend integration, eval pipeline, production logging
Investition
On request
RAG knowledge base
8–12 days. Ingestion pipeline, vector store (Supabase pgvector), retrieval API, golden-set eval, source citations by default
Investition
On request
LLM workflow automation
5–8 days. n8n + LLM steps, 1–3 production workflows, logging and cost caps
Investition
On request
Custom agent
12–20 days. Tool-using agent for a clearly defined domain, eval harness, human-in-the-loop UI, production logging, audit trail
Investition
On request
AI Act disclosure & audit layer
6–10 days. Disclosure UX components, audit logging pipeline, training data documentation, eval pipeline with bias indicators
Investition
On request
Souveränität
Sovereignty:
Jedes Engineering-Paket wird mit EU-Default gebaut. US-Hyperscaler nur bei expliziter Kunden-Entscheidung mit Compliance-Begründung.
- Azure OpenAI in EU region
- Data processing agreement documented
- No training-data sharing
- Optional Mistral / EU models
- Token cost tracking
- Prompt versioning as code
Reifegrad-Map 2026
Use cases by maturity.
What is production-ready in 2026, what works with caveats, and what we don't recommend.
Production-ready
- ·RAG over internal documents (with source citations)
- ·Classification and routing (emails, tickets, documents)
- ·Structured-data extraction from unstructured sources
- ·DE/EN translation, domain-tuned
- ·Code assistance (Cursor, Claude Code, Copilot)
Established patterns, deterministic shares dominate, errors are detectable and correctable.
Recommended with caveats
- ·Multi-step agent workflows (only with clear abort logic and audit trail)
- ·Function calling on production APIs (with idempotency and rollback)
Works, but failure modes are subtle and hard to test. We only build with safeguards.
Not production-ready today
- ·Fully autonomous agents without human in the loop
- ·Creative long-form generation as the main feature (marketing copy, contracts)
Hallucination risk too high, compliance pressure too strong, customer perception too fragile.
EU AI Act · DSGVO
Compliance layer.
EU AI Act and GDPR concretely, not abstractly.
Art. 4 EU AI Act training duty
AI users in the company must be trained. Our training module covers this with attendance proof. (In force.)
Art. 5 EU AI Act prohibited practices
Sprint Zero audit detects violations (social scoring, real-time biometric ID, manipulative systems). (In force.)
Transparency obligations Art. 50 (from August 2026)
Labelling of AI interaction and AI-generated content. Disclosure UX components cover this.
High-risk systems (expected end of 2027)
Postponed by the Digital Omnibus to an expected 2027-12-02 (formal adoption still pending). Preparation window is now. Compliance documentation (risk management, data quality, logging, human oversight) as a separate item, not included in the build sprint.
GDPR data processing
Contract with Microsoft (Azure OpenAI EU) is the standard DPA. No training-data sharing, contractually guaranteed. No US Cloud Act access in EU region.
Wenn KI nicht passt
When AI is not the answer.
Anwendungsfall
Rule-based workflows
Stattdessen
Classic logic. Faster, cheaper, deterministically testable.
Anwendungsfall
Simple CRUD applications
Stattdessen
Standard backend with validation. No LLM needed for input processing.
Anwendungsfall
High-precision calculations (tax, finance)
Stattdessen
Code, not a model. LLMs miscalculate and pretend not to.
Anwendungsfall
Low volume with clear rules
Stattdessen
Manual process plus script. AI setup exceeds the value.
FAQ
FAQ AI Integration.
Can we host on-premise instead of Azure?
Yes, with Mistral or Llama models on your own infrastructure. Higher effort, latency and maintenance on your side. Sensible for strict data classification or air-gap requirements.
How do you prevent hallucinations in production RAG systems?
Source citations by default, confidence thresholds, fallback to 'no answer found', eval suite with golden-set questions. No RAG without an eval pipeline.
Which vector stores do you use?
Default: pgvector in Postgres (EU region). For larger volume: Qdrant self-hosted or Weaviate. Pinecone (US) only by explicit customer decision.
How do you handle multilingual (DE/EN) RAG?
Embedding model with multilingual support (e.g. text-embedding-3-large). Indexed per language separately, query language detected to determine the index. No automatic translation in the index.
What does a typical RAG system cost monthly to run?
Depends heavily on usage volume and model choice. We build token cost tracking and budget limits in by default and deliver a realistic range for your use case in the audit.
AI Engineering im Build Sprint
KI als Engineering-Lieferung. Souverän per Default.
Engineering-Pakete laufen im Rahmen eines Build Sprints — kein separater Audit-Verkauf. Sprint Zero klärt Use-Case, Datenlage, AI-Act-Pflichten und Stack vor dem Build.