One podcast pipeline’s cover art broke when its external image URL died; the fix was routing RSS cover URLs through a single local-file method, plus a couple of quality-of-life tweaks for a language-learning audio feed.
Investing talk kept leaking into work notes and vice versa — the fix was splitting Hermes into a router plus dedicated stock, work, diary, and life sub-assistants, each with its own knowledge base partition and a strict no-cross-referencing prompt.
Podsync started throwing YouTube bot-verification and rate-limit errors — the fix was exporting browser cookies into the container’s mounted data directory and wiring them into yt-dlp’s args.
A personal assistant on the home NAS, a home assistant tied into an investment system, and an intelligence-gathering agent on a VPS — three modules starting to work together as a real perception-cognition-action loop, not just a pile of tools.
A full writeup of the LiteLLM gateway config behind my AI stack — free NVIDIA NIM and OpenRouter models as the safety net, cheap paid models like DeepSeek V3 for daily use, and Claude Sonnet as an on-demand top-tier fallback.
After getting Hermes running on a VPS, I moved it to my Synology NAS instead so it could gradually gain access to household journal and investment data as an internal AI hub — and redesigned the model routing along the way.
OpenClaw’s buzz is fading and everyone’s talking about Hermes — its token compression alone (50%+, reportedly up to 70% in some cases) was reason enough to spin up a VPS and try it.
Setting up Microsoft’s VibeVoice to test audio-to-subtitle generation and text-to-speech: the ASR side (multi-speaker SRT generation) turned out production-ready, while TTS — especially Chinese — wasn’t there yet.
A GPU server configured for Wake-on-LAN would shut down cleanly, then reboot itself ~10 seconds later — the culprit was ACPI wakeup left enabled on the network card, letting ordinary LAN broadcast traffic trigger a wake, not just the magic packet.
When my free Google Cloud credits ran out and Gemini started rate-limiting, I rebuilt OpenClaw’s model chain into a three-tier free-and-cheap failover setup — and rethought which use cases were still worth the tokens.
Adding a quiz question was taking 2-3 minutes and getting slower over time — the fix was switching the data file from a JSON array to NDJSON so new items append in one line instead of a full rewrite.
Adding more API keys didn’t fix the slowdowns — the real issue was a key that existed in secrets but was never registered for failover, plus session pinning that kept using the same key after a restart.
