Custody Map
Every data, model, administrator and support path.
Private AI · Switzerland & Liechtenstein
Private AI that arrives with proof.
Proved before procurement and delivered with the records needed to operate, audit, change, recover and exit. Your archive stays under the custody boundary you choose.
01Custody mapped02Workload frozen03Answers inspectable04Exit recoverable
The differentiator
Five objects replace five promises.
Benchmark Contract
Quality and capacity gates before purchase.
Answer Packet
Sources, claims, authorization and runtime.
Change Receipt
Reason, regression, approval and rollback.
Recovery & Exit
Restore proof and a client-owned rebuild path.
Find the right starting class
Three constraints are enough for a first planning range. The recommendation runs locally and is replaced by measured architecture work before procurement.
Your planning result
Choose the custody, user count and hardest workload.
Indicative CHF bands are dated and non-binding. See inclusions and exclusions →
4custody models
12source-tracked model candidates
7benchmark scenarios
0external runtime requests
The problem, stated quietly
Banks, family offices, trustees, law firms, and clinics hold decades of confidential material: correspondence, contracts, board packs, estate files, client records. Modern AI could make this archive answerable — but not at the price of placing it in uncontrolled external systems.
ANULUM builds the alternative: AI infrastructure that runs where you decide, indexes what you approve, answers with citations from your own documents, and writes an audit trail you can show to whoever holds you accountable.
What we deliver
Local or Swiss-hosted model serving
Language models run on your hardware or on dedicated Swiss infrastructure — never on shared external APIs by default.
Confidential document ingestion
Controlled intake of PDFs, office files, scans, and correspondence, with source hashing and sensitivity classes.
Source-grounded retrieval
Every answer cites the documents it came from. If the archive does not support an answer, the system says so.
Audit logs and version evidence
Prompts, responses, model versions, and index states are recorded — exportable for compliance review.
Evaluation before reliance
Groundedness, refusal behaviour, prompt-injection resistance, and access control are tested before your team relies on the system.
Operation and support
Patching, monitored backups, controlled model updates, and incident response under a defined support agreement.
Deployment on your terms
Four custody models, selected during the architecture review:
01
Air-gapped
No internet path at all. Offline updates. For the most sensitive archives.
02
On-premises
Runs in your building on your network, with controlled maintenance windows.
03
Swiss-hosted dedicated
Single-tenant hardware in a certified Swiss data centre when you cannot host it yourself.
04
Hybrid, policy-gated
Local by default; external models only for approved, non-sensitive tasks — with routing logs.
Who it is for
Family offices
Institutional memory across generations — without exposing the family archive.
Family offices →Trustees & fiduciaries
Source-grounded retrieval across trust deeds, registers, minutes, and client files.
Trustees & fiduciaries →Private banking
Controlled internal AI over policies, research, and regulated knowledge.
Private banking →Research & board offices
Private literature, strategy papers, and board material — searchable, citable, local.
Start with a review →How an engagement works
- Confidential architecture review1–2 weeks, fixed scope We map the archive, the risk boundary, the deployment mode, and the operating cost — before you buy anything.
- Proof node2–6 weeks A limited, sanitised part of the archive runs on real infrastructure. You see citations, refusals, and audit logs on your own material. Then you decide: stop, expand, or productionise.
- Production system and managed operation6–16 weeks, then ongoing Warrantied hardware, hardened deployment, user roles, backups, evaluation pack, training — and a support contract that keeps it dependable.
Inspect the practice before you engage it
Engineering experience
See the public software and systems disciplines behind our private-AI work, with the boundary between public evidence and client delivery stated clearly.
Inspect the engineering →Ten-gate lifecycle
Follow every stage from confidential discovery and proof node through acceptance, controlled change and an exportable exit.
Follow the lifecycle →Practical questions
Get direct answers on data custody, model replacement, concurrency, air-gap updates, compliance boundaries, pricing and ownership.
Read questions and answers →Dense systems produce heat. We design where it goes.
For larger GPU deployments, ANULUM can design liquid-cooling and heat-recovery paths so part of the thermal energy supports building heat or hot-water preheating — engineered with licensed HVAC partners, after a site survey, with measured assumptions.
Expanded programme · 29 July 2026
Three clearer ways to move forward.
Kimi K3 Sovereign Evaluation
Licence, workload, rack feasibility and a measured stop or proceed packet—without a borrowed performance promise.
Open the evaluation →Engagement ladder
Compare discovery, review, pilot, production, operation and exit as separately accepted decisions.
Compare engagements →Trust Centre
Inspect current controls, shared responsibility, precise data-location questions and explicit non-claims.
Inspect the Trust Centre →Named accountability
Founder-led from first decision to recovery drill.
Miroslav Šotek leads ANULUM’s architecture and engineering work in Marbach SG. Public research identity and code history stay inspectable; confidential client environments stay private.