RTX 4090-class GPU
24 GB-class accelerator target for capable local inference and workflow demonstrations.
Proposed prototype · Not yet assembled
MK-II local AI server
The machine intended to prove the idea.
MK-II is the proposed private AI productivity server: a single workstation-class system designed to run useful local models, approved business data, and tailored workflows across a private LAN.
PurposeLocal productivity infrastructure
Email and report drafting, data analysis, private file and database access, and tailored workflows—with people reviewing consequential actions.
Not the focusGeneral appliance control
01 · Hardware target
One capable workstation. Deliberate headroom.
The MK-II target prioritizes a practical single-GPU system that can demonstrate credible local workflows without implying that every future installation needs the same configuration.
128 GB RAM
Headroom for local services, document processing, databases, and CPU-assisted inference.
2 TB NVMe
Operating system, containers, applications, active models, and working data.
4–8 TB
Separate capacity for approved files, databases, vector indexes, backups, and expansion.
2.5 GbE base
Private LAN access for authorized devices, with 10 GbE as a preferred expansion path.
1500–2200 VA UPS
Protected shutdown and short-duration continuity for a workstation-class system.
Power and enclosure planning
Workstation-class airflow, a high-quality power supply sized near 1600 W, a dedicated 120 V circuit where appropriate, and an expected peak system envelope of roughly 800–1200 W would be evaluated during assembly.
02 · Proposed platform
A local stack that can be inspected, replaced, and improved.
The software choices are an initial architecture, not permanent dependencies. The goal is a maintainable private platform with clear boundaries and recoverable data.
Foundation
Ubuntu Server LTS and containerized services managed with Docker Compose.
Local inference
Ollama as the initial local model runtime, with room to evaluate other serving engines.
User workspace
Open WebUI for authorized browser access from devices on the private network.
Workflow layer
n8n for carefully scoped automations, approvals, and repeatable business processes.
Knowledge layer
PostgreSQL and Qdrant for structured records, document retrieval, and cited local search.
Operations
Caddy for local service routing and Restic for encrypted, testable backup routines.
03 · Prototype demonstrations
What MK-II should prove.
A successful prototype is not simply a machine that can produce text. It should complete useful private workflows, show its sources, preserve human review, and reveal the real operational cost.
- 01Draft emails and reports for human review
- 02Analyze approved spreadsheets and business data
- 03Search private files and databases with source citations
- 04Support tailored, repeatable office workflows
- 05Operate across an authorized building LAN
- 06Remain useful when routine public-cloud access is unavailable
04 · Funding objective
Fund measurable work—not a vague promise.
The prototype budget would be tied to a visible sequence of assembly, platform setup, workflow validation, and documented results. The target is proof, not mass production.
Assemble the workstation
Purchase and integrate the compute, memory, storage, networking, cooling, power supply, and UPS within the target envelope.
Establish the private platform
Harden the operating system, deploy the local service stack, define permissions, and confirm backup and recovery behavior.
Prove one complete workflow
Demonstrate an approved document or data workflow from ingestion through cited output and human review.
Measure and document
Record response speed, memory use, power demand, setup time, limitations, and the requirements for a real pilot installation.
Specifications, costs, availability, and software choices remain subject to prototype validation. This page is not a product listing, preorder, investment offer, or performance guarantee.
Help move MK-II forward
A workflow, a pilot conversation, or suitable hardware can all create progress.
Potential pilot users and in-kind hardware partners can contact the project without making a financial commitment.
Contact the project