I build automation pipelines, agent harnesses, and local-first systems where humans stay in the loop—inspectable context, explicit gates, and production-shaped backends (FastAPI, PostgreSQL, Docker, observability).
- Problem: Agent workflows lose intent across sessions; untrusted content reaches LLMs; production ops lack inspectable, human-gated context.
- Solution: Guard–Guide–Build — SCP (input safety), OpenHarness (handoffs + gates), OpenGrimoire (context graph), plus production platform (CaptionPipeline / Platform API).
- Impact: CaptionPipeline: 256+ caption files, 330+ content hours, <1% errors across 9 production feeds (case study); SCP: 16/16 promptfoo tier probes (OWASP LLM01/LLM06); OpenHarness: harness pin-able by commit SHA, autoresearch Tier B 5/5 on foam-pkm + frontend-a2ui skills (Jun 2026).
flowchart TB
Intent[Human / Operator intent]
OH[OpenHarness Guide]
SCP[SCP Guard]
WT[moltbook_watchtower Watch]
AF[Arc_Forge Compounding]
subgraph build [Build]
MOP[media-ops-platform Platform]
OG[OpenGrimoire Context]
end
Intent --> OH
OH --> SCP
SCP -->|gates tools| MOP
SCP -->|gates tools| OG
WT -.->|observe| SCP
AF -.->|mirror| OH
AF -.->|compound| OG
These six repos are the proof set—harness → watch → platform → context → compounding → safety.
| Step | Repo | Command / link |
|---|---|---|
| 1 | OpenHarness | python scripts/verify_script_index.py (from repo root) |
| 2 | SCP | npx promptfoo eval (see README § Testing) |
| 3 | media-ops-platform | README Quick start — API + stack smoke |
| 4 | OpenGrimoire | npm install && npm run dev or CI workflow |
| 5 | Arc_Forge | pytest (workflow_ui suite) |
- CaptionPipeline — automated WhisperX → SCC captions across 9 feeds; Dec 2025 snapshot: 93.5%+ success, peaks 121 files/day → portfolio kit
- SCP guardrail — 16/16 promptfoo injection/reversal probes before LLM context → SCP README § Impact
- Agent harness eval — Tier B 5/5 on foam-pkm and frontend-a2ui skills (2026-06) → OpenHarness + MiscRepos autoresearch harness
- Agent harnesses, handoffs, and intent-alignment gates
- Read-only observability and leak/injection analysis for agent networks
- Local-first context graphs and human↔agent alignment workflows
- LLM input safety (inspect, sanitize, contain, quarantine)
- FastAPI platform APIs, SSO/gateway patterns, homelab observability
| Project | One line | CI |
|---|---|---|
| OpenHarness | Portable harness: context engineering, handoff flow, intent alignment | |
| moltbook_watchtower | Read-only observability for agent networks (leak / injection / behavior) | — |
| media-ops-platform | CaptionPipeline + Platform API — video captions and homelab integration | |
| OpenGrimoire | Local-first context graph and Sync Session alignment workspace | |
| Arc_Forge | Harness mirror + LLM-Wiki compounding in Obsidian | |
| SCP | Secure Contain Protect — MCP guardrail for LLM inputs (OWASP LLM01/LLM06) | |
Portfolio case studies and audit artifacts: media-ops-platform/docs/portfolio/.
Pending
Open an issue on any pinned repo for collaboration or questions.