Open-source AI, built and run on your infrastructure.
We design, build and run AI systems on free and open-source tools and open-weight models. Your data stays in your environment, you are not locked to one vendor, and every component can be inspected.
Four reasons it matters.
- Data control. Models and data can run on your own servers or in your chosen region.
- No lock-in. Components can be swapped as better tools appear, without a rebuild.
- Cost. No per-seat licences for the core stack. You pay for infrastructure and the work.
- Auditability. Every component can be inspected, which matters for compliance and the EU AI Act.
Eight capabilities, one controlled architecture.
We choose tools per project. Whatever the stack, the same rule applies: AI prepares, deterministic checks verify, and a person approves anything consequential.
| Capability | What you get | Tools |
|---|---|---|
| Agent workflows with human approval | Multi-agent workflows with planning, routing and state, where every consequential step passes verification and approval gates before it executes. | LangGraph LangChain CrewAI AutoGen MCP / FastMCP Pydantic AI DSPy Flowise n8n |
| Self-hosted language models | Open-weight models running on your own infrastructure, with routing between local and hosted models by task. | Ollama vLLM LiteLLM Llama Qwen DeepSeek Mistral |
| Document intelligence and OCR | Reading scanned and digital documents, extracting fields and tables, and turning them into structured data. | Tesseract OCR PaddleOCR-VL Surya OCR olmOCR pdfplumber |
| Search and knowledge | Retrieval with citations: vector, keyword and hybrid search, and graph-based retrieval across connected records. | Qdrant ChromaDB FAISS pgvector Elasticsearch OpenSearch BM25 + RRF hybrid search Neo4j (Graph RAG) LlamaIndex Sentence Transformers Playwright Crawl4AI Firecrawl |
| Speech and voice | Transcription, speaker identification, voice detection and speech output for calls, meetings and voice workflows, built by our engineering team and delivery network. | Whisper / faster-whisper Piper TTS Silero VAD SpeechBrain |
| Computer vision | Image-based workflows such as visual inspection, defect and anomaly detection, and photo-based checks, built by our engineering team and delivery network. | OpenCV YOLO DeepLabV3 PatchCore |
| Verification, evaluation and observability | Policy checks before actions, output validation, quality testing of AI answers, and tracing of every model call. | Guardrails AI OPA RAGAS DeepEval Langfuse PyTest Jest |
| Deployment on your infrastructure | Containerised, self-hostable delivery with automated builds and an append-only audit log. | Docker Docker Compose Kubernetes GitHub Actions Jenkins Nginx Linux (RHEL, Ubuntu) NATS Temporal |
Llama, Qwen, DeepSeek and Mistral are open-weight models; their licences differ and are checked per project. n8n is source-available under its Sustainable Use License.
From assessment to a system you run.
| 01 | Assess | Pick one workflow, measure it, and check data, security and hosting constraints. |
| 02 | Choose the stack | Select the tools and models that fit the workflow, and say where no AI is needed. |
| 03 | Build with verification | Every agent action passes checks before it executes. Uncertain cases go to a person. |
| 04 | Deploy on your infrastructure | Self-hosted or in your chosen cloud region, packaged as one repeatable unit. |
| 05 | Monitor and improve | Trace cost, latency and quality, and test answers against agreed standards. |
Public sector: we also assess how our workflows can complement existing open-source public-sector AI platforms. See AI for Government & Public Administration.
The everyday tools behind the systems.
PyTorch TensorFlow Keras scikit-learn Hugging Face Transformers NLTK Gensim NumPy Pandas Matplotlib Seaborn FastAPI Flask Django Django REST Framework AsyncIO Pydantic Express.js NestJS Spring Boot ASP.NET Core Celery Apache Kafka Redis PostgreSQL MySQL SQLite MongoDB React Next.js Vite Streamlit MUI Tailwind CSS Kotlin Jetpack Compose Bash