I’ve spent over 900 hours on Nucleate, but traction has been disappointing—roughly a 1% click-through from Reddit views to the product page. I’ve decided to pivot toward game development, putting deliberate structure behind Jumpsuit (working title: Orbital Jump), though I’m still uncertain how committed I’ll be.

I wrapped up Nucleate packaging—FFmpeg and Ollama installs, clearer Mac error messages for non-admin accounts—and released v1 on Gumroad. I also started Godex, a Qt app using Ollama and embedded Godot docs for local RAG, but hallucinations and my Cursor subscription budget are open risks.

Nucleate’s real strength is integrating with Obsidian and Notion, but I’m still struggling to communicate that. I skipped beta, refactored AppState and Control Center down to ~500 lines with Codex, and upgraded to OpenAI 4.0+ with VRAM-based Qwen 3 versus Mistral selection ahead of a May 15 deadline.

I’ve been trying to get feedback on Nucleate for a while now, but it’s been a mess. I started by reaching out to my brother, then tried public betas, productivity forums, Obsidian/Notion communities, Reddit, Discord groups, and even alumni networks. Obsidian banned me for mentioning the project, saying their…

I’ve been stuck on Nucleate for about a month waiting on feedback from my brother on the beta build after Mac signing and notarization. I’m preparing the Windows package and decided to push release soon even with minimal beta input, while Jumpsuit and Godot learning move slowly around my day job.

I spent the weekend finally getting Nucleate’s Mac app signing and notarization process sorted after months of being blocked by Gatekeeper issues. The app had been failing on new Macs because it was unsigned, and I had to pay for an Apple Developer Program membership (~$99/year) to fix it. Generated signing certificates for the…

I’ve been really focused on security and stability lately, mostly because my brother pushed me to address execution issues and dependency handling. Moving API keys out of plain-text config files into the keyring felt like a step in the right direction, but I’m still not sure how secure that setup is in real-world…

I added information overlays and drag-and-drop audio import, converted the directories panel into a first-start walkthrough, and broadened the Rebuild Summaries panel into a general review summary tool. Mac repackaging stability is still an open question after recent onboarding changes.

A pre-beta test with my wife exposed unclear Nucleate positioning, weak Mac Intel messaging, and broken links—so I integrated GitHub Reporter and a Penny for your thoughts form. I also built a 6DOF Godot ship controller and refactored startup into four interactive cards.

I’ve been working on resolving issues with Obsidian page status indicators in Nucleate, and after packaging it for Mac testing, I sent beta licenses to my brother and wife for their feedback. Codex by OpenAI made around 160 changes in the project, catching several bugs. This experience has left me contemplating…

I reworked the summary backend around progressive compression—summarizing chunks individually before combining—to cut bias and token waste. I overhauled the pipeline into cleanup, chunking, summarization, fact extraction, synthesis, and stylization phases, and raised OpenAI’s default chunk size from 1500 to 6000 tokens.

I refined Nucleate sales page visuals—including a microphone-to-water-to-crystal GIF—and packaged a Mac DMG for pre-beta testers. I fixed an auto notes bug, added a beta checkpoint for summaries, and implemented progressive compression for monthly and weekly transcript summaries.

I tested Nucleate packaging across platforms, with Apple Silicon handling Faster Whisper well since it skips Torch. I stabilized Whisper on Torch 2.8, packaged the Mac app as a .dmg, set up a Windows installer with CUDA support, and fixed Faster Whisper timeouts by switching subprocess logging to debug output.

I’m porting Nucleate from Windows to Mac; Apple Silicon shows promising CPU gains and Metal support is in place. Version 0.5 is official, but packaged cross-platform testing is still pending, and bugs remain around MPS slowness, repeated resummarization, and hub event ordering.

I set up beta screening and feedback forms, added speaker identification as an optional pro toggle, and kept diarization free. I refactored Whisper downloads into subprocesses with a process manager, bought a Mac Mini for packaging tests, and I’m still fighting GUI flicker from full app reboots on Mac.

PyInstaller is working well for Windows CPU-only builds, but Mac Intel still struggles importing PyTorch and pyannote at runtime. I made Faster Whisper the default engine, added stability checkpoints, and plan Mac recording tests on my wife’s Mac while debating free versus pro feature splits ahead of beta.

I focused on Nucleate diarization and chunking—refactoring the split function to break transcripts on reasonable lines while looking back at previous speakers. I also fixed loading bars, app title updates, and smaller transcription pipeline bugs.

Most of the week went into Nucleate on Mac—Intel hardware lacks acceleration, so I’m switching from OpenAI Whisper to Faster Whisper. I also settled on PixelCAD as the Blender pipeline name and added Nucleate menu documentation with panel screenshots while avoiding URL changes that could break app links.

I reorganized Nucleate’s advanced settings, added a maximum recording time feature, and set up a Gumroad publishing page. I still need launch collateral, an animated splash card that doesn’t hurt performance, and a daily-limit alert in the free tier that doesn’t feel like crippleware.

I centralized Nucleate styling into a single theme class, added Ollama detection and an Overdrive toggle, and started diarization work behind pro gating. Mac transcription and summarization remain slower than Windows, and dependency conflicts forced a careful Hugging Face uninstall and modular reinstall.

I added a terminal-style hub window that streams backend actions through a new event bus, plus welcome text to orient users. Faster Whisper downloads now show chunky 5% progress, and I’m still tuning split-flap catchphrase animation fonts on the loading screen.

I refactored the PySide 6 backend so panels refresh after AppState changes—12 of 13 stuck panels now update correctly. I also fixed hub activation race conditions, throttled excessive config refreshes, and kept chipping away at runtime errors in hub animations.

I refactored the app from Tkinter to PySide 6 with the main GUIs operational in the new environment. First-time startup ordering, PySide 6 launch sequencing, and copying files into the user directory on first launch are still the main technical risks.

A week into the Tkinter-to-PySide 6 refactor, I’ve ported panels including System Menu, About, Licensing, Performance, Statistics, Logs, Integrations, File Directory, and AI engine settings. I’m halfway through the AI engine work, with Nucleate Labs still waiting on the remaining panel.

I swapped Tkinter for PySide 6 this week—a significant app design change that brought new challenges but also cleaner structure. I’m working through the migration panel by panel while keeping the tray app and backend behavior intact.

I’ve been working hard on refining the look of Nucleate, focusing on polishing GUIs, refactoring animations, and making them darker and more contrasted. I created a new scroll text class and figured out a solution for the scroll text boxes by hiding the original scroll bar and adding a fake one on top. I’ve completed…

I’ve been redoing crystalline splash animations for Nucleate in Blender using geometry nodes—7 of 10 done, each taking roughly 5 to 8 hours. The complex splash page animation and NLA track editor are still causing slow loads and occasional GUI hangs on larger GIFs.

I improved auto notes with post-processing that culls special notes when content is thin, then spent more time than expected on naming—Echo conflicted with Amazon, so I settled on Nucleate after alternatives like Tiny Logger, Crystallization, and Amber.

I refactored Echo Labs around the mode builder so users can define custom summarization profiles, added daily/weekly/monthly mode tracking, and fixed transcription editors and summarizers that broke in the process. An assistant professor user mode is in as an early experiment.

I refactored the app toward more flexible user modes—OpenAI model selection, daily/weekly/monthly versions, and Ollama backend auto-rebuilds. I’m still weighing Tiny Logger or Logger as a rename (superseded by Nucleate) and monitoring the cost of dropping Metal on Mac.

I fixed animation inconsistencies, normalized the menu and first-time setup flow, and formalized licensing prompts. Removing Torch and MPS dependencies cut the Windows EXE to about 300 MB; I also refactored the mode builder though multi-tag parsing still needs work.

I wrapped a first draft of Echo’s animations but first-time setup bugs slowed progress, including an uninitialized dichromatic theme on the EULA prompt. I simplified the Ollama create function to auto-assign the user’s target model after a Python import refactor caused trouble.

I focused on Echo UI GIF animations in a fluid hand-drawn style, exporting layered PNGs from Inkscape and assembling them in GIMP. I also cleaned up jitter on the language model settings screen and experimented with Blender and SVG imports for a more manipulable pipeline.

I spent the week fixing bugs and refactoring Echo for stability, including a splash screen and GIF logo—though some animations like the atom feel too busy. I also started a color-mapped waveform analyzer for the tray app with a moving read head.

I’ve been diving deep into the Echo app this week, focusing on refactoring the code and configuration for a cleaner, persistent app run. Moved the setup bat script into the main tray app, and it feels good to have everything consolidated in one place.

I expanded Echo user modes from three to about fifteen across sales, medical, marketing, and transcription roles. The config file is near 3,000 lines and reloads every 10 seconds during watch tasks, which is starting to hurt performance.

I redesigned Echo’s icons toward simpler geometric waveforms and refactored the menu around automation plus prototype summarization prompts. I built a special notes manager supporting up to 25 note types and wired it into daily, weekly, and monthly summarizers.

I packaged Echo alpha V00 with a setup wizard, loading bars for model downloads, and early documentation, then launched a refreshed studio website. I’m also refactoring the tray icon busy-state polling, which sometimes fails to update during transcription.

This week’s been pretty focused on refactoring The Echo app, with a heavy emphasis on user experience and new features. I stumbled upon a widget with a background threading issue, so I’ll need to investigate that. I also made a script that detects if an NVIDIA GPU isn’t available and switches to OpenAI, and added a…

I refined Echo and implemented a user-friendly GUI for Notion integration. Whether to add custom profiles and models is still open—that path could mean extensive rewrites if I commit to it.

I added Llama 2 via Ollama and Mistral support to the Scribe app, with a BAT installer that checks Python and Ollama, sets up a venv, and prompts for model selection. Full transcription and summarization testing in a VM is still incomplete.

I wrestled with the studio logo scaling, finished LLC paperwork and the EIN, and evaluated Docker before moving toward PyInstaller packaging instead. I’m laying website groundwork and reworking the first app toward an initial release before diving back into game art.

The LLC paperwork has been approved and corrected. Next steps include filing for an EIN and setting up a separate bank account for the studio.

I’ve made some big changes this week, starting with upgrading my hardware to AMD’s AM5 and a 7900 CPU. The performance boost is impressive—Cinebench scores have nearly doubled, and thermals are well-controlled at just 16 degrees Celsius. It’s a noticeable improvement that’s really helped my workflow.

I’ve been diving deep into Blender AutoTerrain this week, and it’s been a rollercoaster. I’ve noticed some cartoonish effects when I zoom out, which isn’t quite the look I’m aiming for. So, I’ve been tinkering with water ripple effects and exploring the shader editor and geometry nodes.

I’ve been juggling a few things this week. First off, I attended a work conference, but due to some remote setup issues with TeamViewer, I didn’t get much game development work done. So, I’m giving Parsec a try to see if it improves things.

I’ve been tinkering with the grass for the mini scene in the game, and it’s been a bit of a roller coaster ride. I initially ran into issues with grass rendering in pixel art, so I decided to simplify things by reducing the number of grass assets and switching to a 2D polygon asset. However, I encountered another…

I’m experimenting with Blender for environmental 3D work while keeping Aseprite and Unity for hand-drawn sprites. Grid-based path tiling with Bezier noise is promising, but height complexity could explode the tile set from 13 types to over 200.

I settled on Watchlight Studio, secured the domain for five years, and started the Hugo site with Hextra. The local scribe transcription service is buggy and may need a watchdog; I’m also building Blender geometry-node setups for low-poly pixel art rendering.

The local scribe project’s transcription, summarization, and tagging core mostly works, but multi-file analysis still breaks down and single-file summaries run too long. I’m weighing matching interfaces, KeyBert backlinks, and transcript scoring to tighten output.

My local scribe project transcribes audio into Obsidian with Faster Whisper and is moving toward configurable inputs, archiving, and Mistral summarization with a large context window. Notifications currently break batch transcription, so they’re disabled for now.

I’m testing local Whisper transcription for game-dev notes and organizing thoughts, with an eye toward a chat agent and separate local-model collaboration later. The main risk is models losing context across fragmented sessions.