Architecture Atlas

Explain the system without exposing the system.

This page describes functions, responsibilities, and governance boundaries. It is intentionally not a production blueprint.

06

Human authority

Consent, permissions, physical safety, escalation, and final decision rights.

05

Evidence & governance

Audit records, corrections, capability gates, and publication boundaries.

04

Presence & interface

Voice, interaction, and carefully bounded awareness at human scale.

03

Continuity

Memory, identity, history, values, and context that survive a session.

02

Orchestration

Routing, tasks, agents, tools, and the logic that connects capability to purpose.

01

Model layer

Replaceable reasoning engines selected for the work, not treated as the whole identity.

Current hardware backbone

The local machine behind Washu.

This is the primary compute system carrying Washu’s local AI workload today. It is a serious workstation rather than a reference build, assembled to run large open-weight models with enough headroom for memory, perception, voice, agents, and experiments to operate together.

Current configuration · August 2026

GPU compute4× NVIDIA RTX PRO 6000

Blackwell workstation GPUs with 96 GB of VRAM each.

384 GB total VRAM
CPUAMD Threadripper PRO 9975WX

The workstation-class processor at the center of the host.

System memory768 GB

6400 ECC RAM supporting large-model loading, offload, and parallel services.

Power2,800 watts · 220V

A dedicated power foundation sized for sustained multi-GPU work.

Energy supportSolar + home battery

On-site generation and battery capacity help support the project’s continuous local workload.

Scope note: This is the approved public specification of the main workstation. It is not a production topology, access guide, or complete inventory of the home.

WHAT COMES NEXT

From hardware to behavior.

The hardware is only the foundation. Future technical records will document the software stack and then show, through reviewed diagrams, how models, memory, perception, tools, and human approval work together.

WHAT STAYS PRIVATE

The attack surface.

Production addresses, network topology, precise physical layout, live observability, secrets, raw sensing material, household routines, configuration exports, and implementation details that would make intrusion or imitation easier.

The technical record ahead

Three shelves, added when they are ready.

The public technical material will grow in useful, reviewable pieces rather than becoming another maintenance treadmill.

01

Software Stack

The models, memory systems, orchestration, voice, automation, and supporting services that turn the workstation into Washu.

02

System Flow

Charts showing how information and decisions move through Washu, where continuity lives, and where human authority enters.

03

Washu Tools & Tricks

Practical and playful subsystem profiles, including DJ Washu, Draw, CharEmu, and the other tools that grow around the core.

THE DISCLOSURE LINE

Useful detail without a route into the home.

The public record can explain what the system is made of, what it does, and what the work has taught us. The operational details that would expose the family or production environment remain private.

Read the disclosure framework →