LOCAL AI · GOVERNANCE · SECURITY RESEARCH
An ongoing investigation into whether local AI workers can perform bounded tasks, preserve evidence, undergo independent review, and improve over time without controlling their own security boundaries.
The research question: Can a collection of specialized AI workers perform useful work and progressively improve while external systems retain control of permissions, approvals, and deployment?
A locally hosted Qwen model was used in a modular Django development exercise, producing foundational configuration and elements of landing and registration functionality. The result was not a complete production application.
Breaking large assignments into bounded pieces helped make generated work easier to inspect and carry forward without repeatedly sending an entire project into model context. Quantitative improvements have not yet been published.
The lab uses Linux services, Proxmox, managed VLANs, OPNsense firewall controls, PostgreSQL, and Grafana/Loki observability. Qwen runs locally, and the AI environment is designed around restricted communication paths.
Network segmentation, monitoring, and model refusal behavior are not substitutes for application-layer authorization. Exact private topology and operational details are intentionally omitted.
The research pipeline separates work generation, assessment, next-step decisions, and governance. An important design requirement is preserving the original artifact and review evidence across each handoff.
The complete autonomous lifecycle should not be interpreted as fully deployed or independently audited.
The proposed Trainer will diagnose recurring failure modes and support bounded improvement campaigns, including scheduled evaluation of the Manager. A separate trusted controller is intended to verify authority before approved changes reach application code or infrastructure.
Centralized identity services such as FreeIPA and API identities are possible future extensions, not existing deployed controls.
This page is an introduction, not a commitment to a publishing schedule. Feedback will help determine whether future updates focus on technical walkthroughs, governance tests, or reproducible home-lab exercises.
If the project resonates, share which aspect you would want explored next: bounded AI tasks, trust boundaries, observability, or model evaluation and training.
No email subscription or analytics collection is enabled by this page.