I’ve got the core functions of the local language model scribe project mostly working now—it handles transcription, summarization, and tagging for audio files. The main snag is that the language model struggles when trying to analyze multiple files at once. Single file summaries tend to go overboard with examples and sometimes spit out nonsense, so I need to revisit how those summaries are generated to make them more concise and reliable.

I’m thinking about implementing some kind of matching interface to help the model handle multi-file analysis better, but I’m not sure yet how to approach that. Another idea I’m tossing around is scoring transcripts based on length, quality, and progress to tailor the summary sizes, maybe even adding backlinks using KeyBert to connect related summaries. Still on the fence about whether that’s worth the extra complexity.

The tray app is getting some polish—start/stop toggling works and it can run continuously now. I cleaned up the files and Git repo for better organization, and the Git setup feels solid enough to recover from mistakes reliably. I also set up auto routines for daily, weekly, and monthly devlog summaries that are configurable and resettable, which should help keep things organized going forward.

The custom summarizer inside the app is a bit frustrating. Sometimes it requires restarting the whole app to work properly. I added a function to handle different types of work in the custom summarizer, but it still needs optimization to avoid those restarts. That’s on my list to figure out.

I’m also weighing whether to push forward on building a project manager AI agent or just focus on finishing the scribe project first. The scribe app might be ready for a commercial release soon, but I’m not sure what price point would generate enough income to justify the effort.

On a side note, I’m testing Easy Voice Recorder as a possible replacement for Obsidian’s audio logging since it has a tap-to-record feature that could be handy. I haven’t fully evaluated how well it fits long term or how smoothly it would integrate, so that’s still open.

Meanwhile, I’ve made no major progress yet on auto-tiling tooling for Blender, which is supposed to be a semi-procedural terrain generator for level design. It’s still in the early stages.

There are also some IP and ethics concerns around funding since my employer supports starting a sole proprietorship or LLC for game development. I’m looking into business structures and the income requirements for maintaining them, especially since I need to generate income for at least three out of five years.

Overall, I’m juggling a few technical and strategic questions: how to improve multi-file language model performance, how to clean up summaries, whether to add scoring and backlinks, how to optimize the custom summarizer, and what to prioritize next. Plus, the business side is still fuzzy, especially around funding and release plans.

Two-Line Takeaway

Key step forward

The most important forward progress step is getting the core functions of the local language model scribe project mostly working, including transcription, summarization, tagging, and setting up auto routines for devlog summaries.

Significant barrier or risk

The most important barrier or risk is the language model’s struggle with multi-file analysis, leading to unreliable summaries and the need to revisit summary generation methods while also facing uncertainties around business structure, funding, and commercial viability.

Keynotes

Decisions & Observations

• Decided to improve summary generation for single files to make them more concise and reliable.

Issues & Next Steps

• Needs to resolve multi-file language model struggles, possibly by implementing a matching interface for better analysis.
• Will explore scoring transcripts by length, quality, and progress to tailor summary sizes and consider adding backlinks with KeyBert.
• Needs to optimize the custom summarizer to avoid requiring app restarts.
• Deciding whether to prioritize finishing the scribe project for potential commercial release or start building a project manager AI agent.
• Needs to evaluate Easy Voice Recorder as a replacement for Obsidian’s audio logging for long-term fit and integration.
• Investigating business structures (sole proprietorship vs. LLC) and income requirements for funding and legal compliance.

Uncategorized Notes

• No major progress yet on auto-tiling tooling for Blender; remains in early development.
• Set up auto routines for configurable daily, weekly, and monthly devlog summaries to improve organization.

Generated with Nucleate by Watchlight Studio — distilled from full transcription.