About the lab
An independent engineering lab, public by design.
ForgeGuard AI builds and stewards open-source systems for running AI on hardware you control. The public site exists so anyone — an engineer, a collaborator, a hiring committee — can inspect the work directly rather than take claims on faith.
Technical domains
Where the work happens.
Local AI runtimes
Self-hosted speech synthesis and recognition servers with OpenAI-compatible APIs, explicit hardware targets, and health contracts designed for orchestrators.
Model-serving infrastructure
Container and Helm distribution, baked model weights, warmup-aware readiness, bounded concurrency, and GPU telemetry — the operational half of inference.
Agent systems
Infrastructure for agent workloads: the Hermes Agent maintained distribution ships as versioned container images and desktop installers, and the inference servers expose the health, auth, and streaming contracts agents depend on.
Practical tools
Focused engineering utilities that solve one problem well, shipped with the same documentation and status discipline as the larger systems.
Stewardship
Open work with lineage.
Several ForgeGuard projects began as derivatives of excellent upstream work. Stewardship means that lineage is stated plainly: every project page carries an attribution section, every repository carries a NOTICE file, and divergence from upstream is documented rather than obscured. Licenses are respected in both letter and spirit.
The same discipline applies forward: release status, known limitations, and roadmaps are published beside the capabilities, so downstream users can make honest decisions about depending on the work.
Operating principles
How the lab works.
Evidence before claims
Project pages lead with running UI, operational facts, and request contracts. If a capability cannot be traced to code, tests, or release artifacts, it is not claimed — and limitations are documented next to capabilities.
Local-first by default
The lab optimizes for systems that run on hardware you control. Model weights ship inside images; production sites and services avoid third-party runtime dependencies.
Boring reliability
Pinned versions, deterministic builds, reproducible deployment paths, changelogs, and health endpoints that report what is actually true — warming is warming, failure exits non-zero.
Engagement
Paths into the work.
- Evaluate the engineering — start with the engineering profile and the project index.
- Use the systems — every project links its documentation, releases, and deployment paths.
- Contribute or report — issues and pull requests are open on the GitHub organization.
- Talk to the lab — inquiry categories and expectations live on the contact page.