I added support for Llama 2 via Ollama alongside Mistral in the Scribe app, including a 7-bit quantized variant. The setup process now uses a BAT executable that checks for Python and Ollama installations, sets up a virtual environment, manages dependencies, and prompts users to select their model. I tested this setup in a virtual machine, but I haven’t fully verified the transcription and summarization workflows yet, so that’s still on my list.

I refactored and reorganized the scripts to improve usability and maintainability. This included renaming functions, cleaning up error logging, and reorganizing how transcripts are handled for single and weekly summaries. I also added custom summarizer options for daily, weekly, and monthly intervals, which should give users more flexibility. On the API side, I enabled OpenAI API key support so the app can run via GPT, and I integrated API access for transcription as well.

Hardware compatibility remains a tricky area. I developed an automatic hardware detection method, but it currently only works for NVIDIA GPUs. I’m not sure how well the app performs on AMD or other GPUs—Llama models seem to do poorly on AMD GPUs; non-NVIDIA users may face issues. I partially checked Apple Metal support but haven’t fully tested it. For non-NVIDIA users, the app prompts them to adjust default settings, but I still need to verify Faster Whisper’s CPU mode for these cases.

I added integration with Notion’s API and cloud services, but that still needs more testing to be confident it works smoothly. There’s now a stats menu tracking app launches, summaries generated, and minutes of transcribed audio, plus a reset configuration option. I also renamed the app to “Echo by Watchlight Studio” to give it a clearer identity.

Pricing is tentatively set between $25 and $35, but that’s not final. I’m still thinking about marketing strategies, like offering free codes to early users for reviews, but nothing is locked in yet. I’m also debating whether to add a drop-down menu for model selection and a simple text-based export feature—these could improve usability but might not be worth the extra complexity.

Looking ahead, I want to finalize documentation, create custom model files tailored for daily transcription and business use cases, polish the UI, and check compatibility between Notion and Obsidian integrations. I’m also considering how best to refactor the app to support both generic and business summary workflows more cleanly. There are still open questions about verifying the transcription and summarization pipelines fully, expanding API functionality beyond OpenAI and transcription, and making the app fully functional on Windows. Compatibility with non-NVIDIA GPUs remains a big unknown, and I need to figure out whether the tentative pricing and marketing approaches will resonate with users. Overall, there’s progress, but plenty still to test and refine.

Two-Line Takeaway

Key step forward

The most important forward progress step is adding Llama 2 via Ollama alongside Mistral, improved setup automation, API integrations, and enhanced usability features in the Scribe app.

Significant barrier or risk

The most important barrier or risk is the uncertain hardware compatibility, especially with non-NVIDIA GPUs, and the incomplete verification of transcription and summarization workflows.

Keynotes

Decisions & Observations

• Decided to add support for Llama 2 via Ollama (including 7-bit quantized) alongside Mistral in the Scribe app.
• Implemented a BAT executable for streamlined setup, including Python and Ollama checks, virtual environment creation, dependency management, and model selection prompts.
• Added custom summarizer options for daily, weekly, and monthly intervals to increase user flexibility.
• Added stats menu tracking app usage and a reset configuration option.
• Tentatively set pricing between $25 and $35; marketing strategies (e.g., free codes for early reviews) are under consideration but not finalized.
• Debating adding a drop-down menu for model selection and a text-based export feature to improve usability versus added complexity.

Issues & Next Steps

• Developed automatic hardware detection method, currently limited to NVIDIA GPUs; need to resolve compatibility issues with non-NVIDIA GPUs (AMD, Intel, Apple) support.
• Non-NVIDIA users are prompted to adjust settings; need to verify Faster Whisper’s CPU mode effectiveness for these users.
• Needs to fully verify transcription and summarization workflows.
• Will explore expanding API functionality beyond OpenAI and transcription.
• Plans to finalize documentation, create custom model files for daily transcription and business use cases, polish UI, and check Notion-Obsidian integration compatibility.
• Considering refactoring app architecture to better support both generic and business summary workflows.
• Needs to validate pricing and marketing approaches with user feedback.

Uncategorized Notes

• Refactored and reorganized scripts for better usability and maintainability, including function renaming, error logging cleanup, and transcript handling improvements.
• Enabled OpenAI API key support and integrated API access for transcription.
• Integrated Notion API and cloud services but require further testing to ensure smooth operation.
• Renamed app to “Echo by Watchlight Studio” for clearer branding.
• Must address full Windows functionality and hardware compatibility beyond NVIDIA GPUs.

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