View the source
Follow on LinkedIn
On-premise by design
Runs inside your own walls. Documents, vectors, and conversations stay on hardware you control.
Open source · sovereign AI
Applied AI that never leaves your walls.
ArcumAI is an open-source agentic platform that runs entirely on your own hardware. It reads your documents, answers in context, and keeps every byte inside your infrastructure. No external API. No data exit. No compromise.
- Swiss banking
- Fiduciaries
- Legal offices
- Regulated data residency
// the trade-off
Most AI asks you to
export your secrets.
Mainstream assistants only work by shipping your prompts and your documents to someone else's servers. For a private bank, a fiduciary, or a law firm, that single condition rules them out before the conversation starts.
So ArcumAI inverts the arrangement: the model comes to the data, never the other way around.
// what it does
Built for places where
data must stay local and sovereign.
Local LLM inference
Open models served through Ollama. It runs on a modest server, with no licence keys phoning home.
Hybrid retrieval
Semantic search and keyword search, fused. Answers are grounded in your documents, with sources you can trace.
Agentic, with MCP
Tools and context connect through the Model Context Protocol, so the assistant can act, not just chat.
Isolated cloud option
Need heavier reasoning? An optional cloud path stays hard-isolated from your local data. Off by default.
Open source, MIT
Every line is readable, auditable, and yours to fork. No black box, no vendor lock-in.
// how it works
A retrieval pipeline
that stays on-premise.
- 01
Ingest
Your files are read and chunked locally, then indexed. Nothing is uploaded anywhere.
- 02
Index
Chunks land in a local vector store and a keyword index, side by side.
- 03
Retrieve
A query fuses semantic and keyword results to find the passages that actually matter.
- 04
Reason
A local model reads the retrieved context and composes an answer grounded in it.
- 05
Respond
You get a cited, context-aware reply in the web app or directly inside Outlook.
Interfaces: a NiceGUI web workspace and a native Outlook add-in. Backend: FastAPI. Optional isolated cloud reasoning when you ask for it, never by default.
// under the hood
An open stack,
fully inspectable.
No proprietary runtime sits between you and your data. The pieces are standard, swappable, and documented in the repository.
Ollama
Llama 3.2
LlamaIndex
Apertus LLM
ChromaDB
BM25 / QueryFusion
FastAPI
NiceGUI
MCP
Outlook add-in (C# VSTO)
Data layer: ChromaDB + JSON, no external database
// where it fits
Designed for regulated Switzerland,
and anywhere data sovereignty is essential.
Private banking
Surface client policy, KYC files, and product documentation without a single record crossing the institution's perimeter.
Fiduciaries
Search contracts, mandates, and correspondence in plain language, with answers that stay on the office network.
Legal offices
Question case files and precedent locally, preserving privilege and client confidentiality end to end.
Self-host ArcumAI today.
Clone it, read every line, and run it on your own terms. Open source and free to deploy.