← Selected work
Local AI systems 2026

Nexus / Hermes

A private AI lab where models run on owned hardware, used to test what agents can safely do on their own and where that line should sit.

Active laboratory

The idea

The experiment is moving from isolated prompts toward an environment that can preserve context, assign work, expose state, accept correction, and keep a human approval boundary where it matters.
Nexus / Hermes — public system view
01 Capture Preserve the source material
02 Distill Transform without erasing origin
03 Assign Route bounded work to agents
04 Inspect Expose state, artifacts, and drift
05 Approve Keep consequential writes human-led

Agents need an operating environment

Nexus is where local model serving, retrieval, transcription, task routing, and named agent roles meet. Scout, Librarian, and Forge separate responsibilities so work is easier to inspect, constrain, and hand off.

Memory without flattening the person

A recurring design question is how an assistant can learn context and preferences while remaining its own tool—and while preserving source material, provenance, correction, and user control. The knowledge workflow is treated as infrastructure with explicit sources and boundaries.

Useful before autonomous

The lab favors small, observable workflows over broad claims of autonomy: capture a conversation, preserve the raw record, create a restrained draft, expose the transformation, and require approval before it enters the system of record.

What this work demonstrates

Evidence, not a capability list.

  • A cold-start defect made explicit.

    A path mismatch in the orchestrator's file-write tool was traced to how the harness resolved its workspace and recorded as a known operating boundary.

  • Model choice is measured.

    Models are compared on the same RTX 3090 for memory residency, throughput, and task quality. Capability on paper does not decide what earns a place in the lab.

  • A gateway between agents and the internet.

    External search crosses a private, allowlisted gateway with operator status and a kill switch. The gateway was commissioned and frozen before agents could depend on it.

  • Roles with real permissions.

    Named agent profiles hold different access. Research and drafting can be delegated while consequential writes remain outside those roles.

  • A human approval boundary.

    Consequential writes stop and wait. Human approval is a permanent design position for the system.

Public boundary

This is an active research environment, not a finished product or a claim of production autonomy. Public material focuses on design questions and operating principles.