Edd Mann
@eddmann.com
November 12, 2025 at 9:37 AM
it has been a lot of fun exploring local inference and how you can do this within a swift application.
whisperkit and mlx make it surprisingly easy to run powerful models on apple silicon without any network calls.
whisperkit and mlx make it surprisingly easy to run powerful models on apple silicon without any network calls.
November 12, 2025 at 9:37 AM
it has been a lot of fun exploring local inference and how you can do this within a swift application.
whisperkit and mlx make it surprisingly easy to run powerful models on apple silicon without any network calls.
whisperkit and mlx make it surprisingly easy to run powerful models on apple silicon without any network calls.
the workflow is simple:
- global hotkey (option-shift-space)
- press space to record
- transcription auto-copies and can auto-paste
- recent history stored locally in swiftdata
- global hotkey (option-shift-space)
- press space to record
- transcription auto-copies and can auto-paste
- recent history stored locally in swiftdata
November 12, 2025 at 9:37 AM
the workflow is simple:
- global hotkey (option-shift-space)
- press space to record
- transcription auto-copies and can auto-paste
- recent history stored locally in swiftdata
- global hotkey (option-shift-space)
- press space to record
- transcription auto-copies and can auto-paste
- recent history stored locally in swiftdata
it also gives you the option to use openai's transcription service and gpt-4o-mini for post-processing if you prefer that route.
flexibility is nice - some people want privacy, others want speed/convenience.
flexibility is nice - some people want privacy, others want speed/convenience.
November 12, 2025 at 9:36 AM
it also gives you the option to use openai's transcription service and gpt-4o-mini for post-processing if you prefer that route.
flexibility is nice - some people want privacy, others want speed/convenience.
flexibility is nice - some people want privacy, others want speed/convenience.
the core is pretty straightforward:
- whisperkit for transcription (base/small/medium models)
- mlx for ai post-processing (qwen 2.5, llama 3.2, phi-3.5)
- all models downloaded from hugging face
- all inference happens locally on your mac
- whisperkit for transcription (base/small/medium models)
- mlx for ai post-processing (qwen 2.5, llama 3.2, phi-3.5)
- all models downloaded from hugging face
- all inference happens locally on your mac
November 12, 2025 at 9:36 AM
the core is pretty straightforward:
- whisperkit for transcription (base/small/medium models)
- mlx for ai post-processing (qwen 2.5, llama 3.2, phi-3.5)
- all models downloaded from hugging face
- all inference happens locally on your mac
- whisperkit for transcription (base/small/medium models)
- mlx for ai post-processing (qwen 2.5, llama 3.2, phi-3.5)
- all models downloaded from hugging face
- all inference happens locally on your mac
but i wanted to explore local models - specifically whisperkit and mlx.
so i thought it would be fun to build out my own version that provides 100% local transcription and post-processing, all running on-device.
so i thought it would be fun to build out my own version that provides 100% local transcription and post-processing, all running on-device.
November 12, 2025 at 9:36 AM
but i wanted to explore local models - specifically whisperkit and mlx.
so i thought it would be fun to build out my own version that provides 100% local transcription and post-processing, all running on-device.
so i thought it would be fun to build out my own version that provides 100% local transcription and post-processing, all running on-device.
recently i've been exploring building out some small macOS utilities to help with daily tasks.
i really like the idea of talking more than typing; projects like superwhisper are really cool.
i really like the idea of talking more than typing; projects like superwhisper are really cool.
November 12, 2025 at 9:36 AM
recently i've been exploring building out some small macOS utilities to help with daily tasks.
i really like the idea of talking more than typing; projects like superwhisper are really cool.
i really like the idea of talking more than typing; projects like superwhisper are really cool.
you can check out the final build + full git history from the weekend here:
🔗 github.com/eddmann/phpboy
🔗 github.com/eddmann/phpboy
November 11, 2025 at 1:19 PM
you can check out the final build + full git history from the weekend here:
🔗 github.com/eddmann/phpboy
🔗 github.com/eddmann/phpboy
honestly, i didn’t know what to expect from the remote agent workflow - and doing all this from my phone felt unreal (and a bit of a gimmick) at first.
but by the end of the weekend, it actually started to click. 💭
but by the end of the weekend, it actually started to click. 💭
November 11, 2025 at 1:19 PM
honestly, i didn’t know what to expect from the remote agent workflow - and doing all this from my phone felt unreal (and a bit of a gimmick) at first.
but by the end of the weekend, it actually started to click. 💭
but by the end of the weekend, it actually started to click. 💭
i wanted multiple frontends, including a php-wasm build (which technically works… just incredibly slow). 🐢
with claude’s help, i’ve now ticketed up the remaining issues to work through this week.
with claude’s help, i’ve now ticketed up the remaining issues to work through this week.
November 11, 2025 at 1:19 PM
i wanted multiple frontends, including a php-wasm build (which technically works… just incredibly slow). 🐢
with claude’s help, i’ve now ticketed up the remaining issues to work through this week.
with claude’s help, i’ve now ticketed up the remaining issues to work through this week.
it’s not fully complete yet, a few things still need work:
- full colour support
- audio issues
- frontend: sdl2 desktop build
- frontend: wasm performance
- full colour support
- audio issues
- frontend: sdl2 desktop build
- frontend: wasm performance
November 11, 2025 at 1:19 PM
it’s not fully complete yet, a few things still need work:
- full colour support
- audio issues
- frontend: sdl2 desktop build
- frontend: wasm performance
- full colour support
- audio issues
- frontend: sdl2 desktop build
- frontend: wasm performance
by monday morning, i pulled the repo from github… and ran pokemon red in my terminal. 🤯
November 11, 2025 at 1:19 PM
by monday morning, i pulled the repo from github… and ran pokemon red in my terminal. 🤯
most of my time was spent reviewing the generated test coverage describing the emulator’s behaviour - rather than micromanaging the implementation details.
but throughout the weekend i ended up learning so much - t-cycles vs m-cycles, z80 quirks, apus, vblank timing…
but throughout the weekend i ended up learning so much - t-cycles vs m-cycles, z80 quirks, apus, vblank timing…
November 11, 2025 at 1:19 PM
most of my time was spent reviewing the generated test coverage describing the emulator’s behaviour - rather than micromanaging the implementation details.
but throughout the weekend i ended up learning so much - t-cycles vs m-cycles, z80 quirks, apus, vblank timing…
but throughout the weekend i ended up learning so much - t-cycles vs m-cycles, z80 quirks, apus, vblank timing…
some of the agent’s work i rejected at the PR stage, kicking off slightly different prompt variations to explore multiple approaches before settling on one i liked.
to manage context i would split up deep-dive and implementation work, commiting findings in-between.
to manage context i would split up deep-dive and implementation work, commiting findings in-between.
November 11, 2025 at 1:18 PM
some of the agent’s work i rejected at the PR stage, kicking off slightly different prompt variations to explore multiple approaches before settling on one i liked.
to manage context i would split up deep-dive and implementation work, commiting findings in-between.
to manage context i would split up deep-dive and implementation work, commiting findings in-between.
on saturday afternoon i was literally at a social gathering, sneaking a quick merge to add apu support into the emulator 🤯
it got a bit addictive ngl.
it got a bit addictive ngl.
November 11, 2025 at 1:18 PM
on saturday afternoon i was literally at a social gathering, sneaking a quick merge to add apu support into the emulator 🤯
it got a bit addictive ngl.
it got a bit addictive ngl.
what made it interesting though was the process.
the back-and-forth with agents, github, and ci/cd - all happening from my phone. 📱
the back-and-forth with agents, github, and ci/cd - all happening from my phone. 📱
November 11, 2025 at 1:18 PM
what made it interesting though was the process.
the back-and-forth with agents, github, and ci/cd - all happening from my phone. 📱
the back-and-forth with agents, github, and ci/cd - all happening from my phone. 📱
this isn’t exactly a novel problem space - there’s plenty of prior art, great docs (pandocs), solid reference builds like sameboy, and test roms (blargg, mooneye) to validate correctness.
November 11, 2025 at 1:18 PM
this isn’t exactly a novel problem space - there’s plenty of prior art, great docs (pandocs), solid reference builds like sameboy, and test roms (blargg, mooneye) to validate correctness.
friday night, i put together a PLAN md with cGPT’s help, then got the remote agent to start building step-by-step - each stage had clear goals and outcomes before moving on.
November 11, 2025 at 1:18 PM
friday night, i put together a PLAN md with cGPT’s help, then got the remote agent to start building step-by-step - each stage had clear goals and outcomes before moving on.
first step - come up with a name.
then, obviously, get cGPT to generate some artwork (because why not).
then, obviously, get cGPT to generate some artwork (because why not).
November 11, 2025 at 1:18 PM
first step - come up with a name.
then, obviously, get cGPT to generate some artwork (because why not).
then, obviously, get cGPT to generate some artwork (because why not).
i’ve always wanted to build a game boy emulator.
it’s been a while since i did a personal project in php, so when claude web dropped $1000 of credit to use before nov 19th, i figured it was the perfect excuse to finally build one. 🎮
it’s been a while since i did a personal project in php, so when claude web dropped $1000 of credit to use before nov 19th, i figured it was the perfect excuse to finally build one. 🎮
November 11, 2025 at 1:18 PM
i’ve always wanted to build a game boy emulator.
it’s been a while since i did a personal project in php, so when claude web dropped $1000 of credit to use before nov 19th, i figured it was the perfect excuse to finally build one. 🎮
it’s been a while since i did a personal project in php, so when claude web dropped $1000 of credit to use before nov 19th, i figured it was the perfect excuse to finally build one. 🎮