SYSTEMS

Things that are actually running.

A systems index for the notebook. Local agents, automation workflows, memory, game server operations, and a few practical builds around FPV and visual work.

Active
6 systems
Completed
2 systems
Bias
Useful over shiny
01 — AI / OPS
01
Active

Multi-agent AI organisation

A production multi-agent system running research, code review, content, bots, and monitoring on local infrastructure. As of September 2026 it runs on Hermes: Reidar is the gateway agent, with five bounded bot profiles under it. The Agno version in the March note is where it started.

Research pipeline: 30 min to 50 sec. 8 agents, local models, one home server.
Problem
One model asked to be researcher, coder, reviewer, and manager at once. On complex work the quality falls off a cliff.
Constraint
Local models on one desktop GPU (RTX 5090, 32 GB VRAM). Only the orchestrator uses a cloud API.
What I did
March 2026 design, since moved to Hermes: an org chart. Each agent gets a defined role, a model suited to that role, and a clear reporting line. Code goes through a 4-step pipeline with a separate QA agent. Read note →
Result
Research that used to take 30 minutes takes 50 seconds. A separate model with no context of the original intent catches failure modes the author model misses.
HermesOllamaPythonRTX 5090
Lane
AI · Infra
Runtime
24/7
02
Active

Hermes operator stack

Daily operator surface for site checks, repo inspection, a read-only review lane (Mads, the Claude review agent on OpenClaw until August 2026; the judge bot on Hermes holds that lane now), browser verification, persistent memory, and production follow-through.

One working path for advice first, mutation second, verification last.
HermesReview laneHindsightVerification
Lane
Ops · AI
03
Active

Agent run rubrics

A small local review layer for Hermes agent runs. It checks agent runs against artifacts instead of trusting green prose.

Success is not enough. The useful question is what was actually verified.
Problem
Agent runs report success while tests fail, the output is plan-only, or the state is stale.
Constraint
Deliberately mechanical. No model calls, no automatic memory writes, no deployment decisions, no merge approval.
What I did
A rubric verdict with hard gates per run: was result.json read, did tests fail, do the changed files match the task, was the run review-only or implementation-ready.
Result
A finished run is not mistaken for verified work. The case for it: 41 of 41 suites reported passing while three were failing →
HermesRubricsVerification
Lane
AI · Ops
04
Completed

Claude Design workflow

Repo-grounded frontend generation with no-apply-by-default safety, static verification, scorecards, and manual winner bundles.

Generation creates artifacts. Operators decide what ships.
Claude CodeScorecardsStatic verificationNo-apply
Lane
AI · Frontend
05
Active

Hindsight local memory

Persistent memory routed through a local OpenAI-compatible Ollama endpoint. Built around durable facts rather than dumping every session into prompts.

Memory is useful when it reduces repeated steering without becoming a second, fuzzier source of truth.
HindsightOllamaLocal embeddings
Lane
Memory · Ops
02 — PRACTICAL BUILDS
06
Active

Windrose game server operations

A Windows 11 VM-hosted early-access game server that went live the same hour a friend asked for it, then got stabilised with the AI operations stack around it.

Observed: stable sessions with 8 players on a build other hosts reported trouble with above 4.
Windows 11 VMGame serverAI opsMonitoring
Lane
Gaming · Ops
07
Active

Discord shopping list bot

Discord-native shopping list with quick-add, store-aware sorting, and daily price alerts. Flask plus Nextcloud. Used daily.

PythonDiscord APINextcloud
Lane
Automation
Use
Family · daily
08
Completed

AI-designed 3D printed camera mounts

AI-assisted modelling and print iteration for FPV camera mounts.

BlenderBambu Lab P1SFPV
Lane
FPV · 3D print
Stage
Done