ARTIFICIAL INTELLIGENCE.
Artificial Intelligence is no longer an experimental technology. The challenge is not whether AI can generate answers, but whether it can operate reliably, securely, and predictably inside real organizations.
What we do.
-
We design the target architecture before a model is chosen: business, application and technology layers, a control-plane / compute-plane split, and contracts that keep models replaceable.
-
AI systems that run entirely inside your network — local inference, local embeddings, no document content or prompts sent to external APIs.
-
Grounded question-answering over your own documents: ingestion, chunking, hybrid retrieval, reranking, and answers that cite their sources or decline.
-
Access models, document lifecycle, audit logging, evaluation gates and kill switches, with EU AI Act and GDPR alignment for regulated and healthcare settings.
-
Connecting AI into existing applications, databases and processes with n8n and custom services — idempotent, staged, fail-closed.
-
Executive guidance on where AI fits, build-vs-buy, model selection, and a phased roadmap you can actually deliver.
MEMBER of.
MEMBER of.
How we work.
-
Business, application and technology layers separated (ArchiMate); a control plane for orchestration and data, a compute plane for inference, joined only by stable OpenAI-compatible contracts so models stay replaceable.
-
Canonical terminology, a document lifecycle, an indexing policy and an access model are written before the first document is ingested.
-
Each phase has a dry run, a pre-change backup with checksums, an apply step, and a written report. Decisions are recorded as ADRs.
-
A real evaluation set (30+ graded questions) gates every retrieval or model change. Retrieval hit-rate, latency and refusal behavior are tracked across versions.
-
Invalid input is rejected, not guessed at. Low-confidence answers are flagged for human review. There is a kill switch at UI, orchestration, inference and database level.
Reference build.
private clinical AI environment for a hospital
A fully on-premise AI platform for a hospital: staff chat, a clinical-assistant model, and a retrieval system that answers from the organization's own knowledge base with citations. No data leaves the hospital network — no external LLM API, no internet retrieval.
Split control plane (OpenWebUI · middleware · PostgreSQL + pgvector · n8n) and GPU compute plane (local Ollama serving Mistral, MedGemma and Qwen), connected by OpenAI-compatible APIs.
Custom indexing pipeline: source snapshots → validation → chunking → local embeddings (bge-m3, 1024-dim) → pgvector. 570+ documents, 1,500+ chunks in the first corpus.
Hybrid retrieval (vector + keyword) with reranking and mandatory source citations.
Built to EU AI Act and GDPR requirements: all personal and clinical data is processed on-premise and never leaves the hospital network. The assistant is advisory and physician-in-the-loop — it does not generate diagnoses or treatment recommendations — and therefore falls outside the scope of the Medical Device Regulation.
Delivered components.
-
Private ai assistant
Local LLM chat for company staff through a single web interface — no RAG, no external API.
-
Clinical assistant model
Physician-facing model served locally for clinical Q&A and summarization; physician-in-the-loop.
-
Knowledge-base RAG
Grounded answers over internal IT and operations documentation, with citations and a "don't know" fallback.
-
Secure document ingestion
Converts office files and scans to clean Markdown — deterministic extraction, local OCR only, idempotent, fully audited.
Frameworks & Technologies.
