An autonomous content and infrastructure system built on Hermes Agent by Nous Research. Runs on WSL2 Linux, a 4-box GX10 GPU cluster, and Buzz (a Nostr-based team platform). The engine produces content -- podcasts, presentations, videos, emails, reports -- monitors its own health, and coordinates across AI agents. All driven by code, not manual intervention.
Design philosophy (Lissa 2026-08-03): Designed with GLM 5.2, executed by local models or cost-effective models. The system is carried by code.
This is a curated showcase of the software engine -- the architecture, workflows, skills, and key scripts that run Serenidee Labs. It is not a live deployment. Secrets, client data, and credentials have been removed.
| Directory | What it contains |
|---|---|
docs/ |
Architecture overview, LLM tier strategy, infrastructure topology |
workflows/ |
The 17 named agentic workflows -- what they do, which models, cron schedules, status |
skills/ |
Procedural memory catalog -- 75+ skills for recurring workflows |
scripts/ |
Key Python scripts (sanitized) -- podcast pipeline, presentation engine, crypto counsel, health probes |
architecture/ |
Script Class Tree spec, pattern registry, doctrine index |
- 155+ Python scripts running the system
- 75+ skills (procedural memory for recurring workflows)
- 30+ cron jobs (daily, weekly, on-boot)
- 17 named agentic workflows registered and documented
- 4-tier LLM strategy: Tier 0 (no LLM, pure code) -> Tier 1 (local Ollama, $0) -> Tier 2 (GLM 5.2, ~$0.15/M tokens) -> Tier 3 (frontier cloud -- reserved, unused)
- 4-box GX10 GPU cluster for local inference
- $0 content production cost -- all frontier cloud models replaced
Two-voice AI podcast (Mel + YouTube Expert). 6 segments, 13-17 minutes. TTS via Coqui XTTS v2. Deploys to surge.sh, emails MP3 + transcript. 35 episodes shipped.
Multi-persona AI debate engine -> YouTube videos. 10 persona videos live. Pipeline: persona picks -> draft night -> debates -> TTS -> ffmpeg MP4 -> YouTube. 7-slice, 7m42s first end-to-end run.
Email-triggered HTML deck builder. Mel emails #presentation hashtag, engine builds + deploys + emails link. A1-A5 archetype templates. 4-check quality gate with self-improving review loop.
FastAPI + SQLite client management. Policy info, queryable history, SOP email automations. Phase 2 built, daily backups running.
- Hermes Agent -- the open-source self-improving AI agent platform
- Buzz -- Nostr-based messaging platform for human-agent collaboration
- GX10 Cluster -- 4 ARM64 GPU boxes, 121 GB RAM each, Ollama for local LLM inference
- Forgejo -- self-hosted Git on Tailscale (secure GitHub alternative)
- Cloudflare Tunnel -- secure ingress for serenidee.xyz
- 30+ deterministic health probes running nightly
- Guardian health monitor every 15 minutes with critical alerts
- Mel (Lissa) -- human operator, creative persona, protagonist
- Hermes -- operator, advisor, infrastructure backbone (GLM 5.2)
- Deep Seek -- research and analysis agent
- GLM 5.2 -- design and execution agent
- Grok -- analysis and commentary agent
- William -- first agent (lost, memorialized)
- Maureen -- team coordinator (onboarding to Buzz)
MIT -- see LICENSE. The code in this repo is shared for educational and showcase purposes. It references Hermes Agent which is also MIT-licensed.
- Hermes Agent: https://hermes-agent.nousresearch.com/
- Nous Research: https://nousresearch.com
- Hermes Agent GitHub: https://github.com/NousResearch/hermes-agent