Auditable AI infrastructure · control planes · audit gates · Blackwell kernel work

I build auditable AI infrastructure for autonomous AI systems.

Control planes, audit gates, and kernel-level debugging for GPU work. My system's own strict audit passes 389 of 389 canonical outputs and headlines that gate on the front page. Auditing AI output is the product; the papers are what it audits.

View Enoch View corpus View promising signals Read docs GitHub profile
Generated hero graphic for Enoch showing an agentic research control plane coordinating worker nodes and artifact bundles
389canonical AI-generated artifacts indexed
389/389pass packaging/provenance lint
389/389pass strict claim/evidence audit

Featured project

Enoch: Agentic Research Control Plane

Enoch is an open-source control plane for running bounded AI research workflows end to end. It coordinates idea intake, queue state, pause and maintenance controls, worker preflight, single-lane safety, evidence synchronization, dashboard status, alerting, and publication-style artifact packaging. Start with the local proof, then inspect both the packaging/provenance lint and the strict claim/evidence audit status.

The project is built on FastAPI with a hard state contract and operated with Codex plus oh-my-codex/OMX-assisted development workflows.

System repository → Research corpus → Promising signals → Documentation website → Enoch launch site →

Architecture

From idea queue to auditable artifact.

Enoch is framed as infrastructure first: state, safety, evidence, and review boundaries before generated output.

Enoch architecture diagram showing idea intake, control plane, wake gate, worker, evidence, corpus, dashboard, and alerts

What I care about

  • Agent orchestration with durable state
  • Local AI infrastructure and worker safety
  • Evidence-grounded automation
  • Queue systems that fail loudly instead of hanging silently
  • Human-visible provenance for generated artifacts

Current stack

  • Python, FastAPI, SQLite-backed control state
  • FastAPI control-plane boundaries with a hard state contract
  • GitHub Actions, branch protections, release packaging
  • Wake-gated local worker execution
  • Codex-assisted development and oh-my-codex/OMX operations

Provenance matters

The generated papers in the Enoch corpus are released as AI-generated research artifacts. Separately, the promising-signals repo preserves bounded useful or scale-blocked leads that are not validated papers and not counted as corpus papers. I am not claiming human authorship of the paper prose, arguments, or generated results. The packaging/provenance lint checks required metadata and overclaim patterns; it does not mean the artifacts are peer-reviewed, scientifically correct, independently replicated, or deeply claim/evidence auditable. Current strict claim/evidence audit status is 389/389 passing. The release is about the surrounding system: control-plane design, workflow reliability, evidence capture, provenance metadata, and public packaging.