I built an AI gf companion for a user on the forum for free

You niggas won't believe how much I defend this forum on reddit and other forums then I see shit like this and wonder if those redditors are right
A good chunk of these users are simps and would gladly help a foid infiltrate if it was for a crumb of attention :feelskek:
 
  • +1
Reactions: strongmtn, Chadeep, Jason Voorhees and 1 other person
Fuck yes bro this brought a tear to my eye, will read the technical parts later and try to figure it out bc it seems fun
 
  • +1
Reactions: Chadeep and Jason Voorhees
Can you make me a fleshlight
 
  • +1
Reactions: filthycurrycel, Chadeep and Jason Voorhees
@Divineincel thoughts since you are a technical guy too
 
  • +1
Reactions: filthycurrycel
@81xa @Histy
 
  • +1
  • JFL
Reactions: Histy, filthycurrycel and 81xa
Giving back to the community
 
  • +1
Reactions: Jason Voorhees and filthycurrycel
1781540986183


@jzo @Hess @buccalfatremoval @chang cypionate @primal_shitmuncher
 
  • JFL
  • +1
Reactions: primal_shitmuncher, buccalfatremoval, Chadeep and 3 others
@Wicket @optimisticzoomer @FUCK YEA
 
  • Love it
Reactions: FUCK YEA
@BeanCelll @alurmo @LXR @Notcel @inyerta
 
  • +1
Reactions: LXR
@Former Shortcel @iblamexyz @Final Fantasy
 
  • +1
Reactions: iblamexyz
more than @chance 's internship
What's more insane is suddenly a bunch of niggas got banned just like that like Thanos snapping finger with one single thread fuck
 
  • JFL
Reactions: Subhuman and Glorious King
A pretty well known user came to me wanting a local AI companion app. In all my years when I used to do freelancing. I've never gotten request like that. I've shipped plenty of AI apps since before this stuff was mainstream even tho this isnt even my domain. im a devops guys but I never got a request for something this like this. Something this personal. Took it purely out of curiosity. Ended up having a lot of fun

View attachment 5222869View attachment 5222878View attachment 5223065


His actual requirements were modest. a few hours of work, honestly But that's not my thing. I'm all or nothing kind of guy either I don't touch it or I do it to the absolute best I'm capable of It's why freelance clients kept coming back. I over deliver by default. So instead I spent days building the most over engineered version of this thing I could. Here are the video of it working irl. On my macbook completely but still smooth as hell. Recorded on phone to prove its not editing or some shit. I couldn't upload the deep emotional conversations because of the 10 MB video limit


View attachment 5222887
View attachment 5222888

I know you niggas don't care about the specifics or technical details but I still wrote down my experience building it and shipping it. It's a habit of mine so I'm doing it to scratch that itch and to come back to it future. It's an intriguing read anyway but I've put it in spoilers so you dont have to scroll to much


His machine is an RTX 2080 Super.

View attachment 5222922

Has like 8GB VRAM like a decade old. My first reaction was damn it but tbh anyone can make a model scream on a 5090, the real engineering prowess is making it genuinely good on hardware that old. So I treated the 8GB ceiling as the whole challenge I developed it on my $10K mac studio and dgx spark

but optimized everything heavily for CUDA using Nvidia. The demo above was shot on an M4 MacBook Air and it holds up.

Everyone else shipping stuff like this just dumps an 8B model on hugging face behind a silly tavern ui and call it a day done.
View attachment 5222897
Its power user stack. Tabs, presets, sliders, config files, it looks like a aeroplane cockpit not an ai companion. If I'd handed him that. It would do disservice to him because he'd be doing the technical work wiring a backend, importing cards, fiddling with samplers just to say hello and when you ship something you make it as seamless, as painless and engaging as possible. Good enough, it works is not enough for me. So I built everything from scratch. Custom frontend, an LLM orchestration layers, a proper personality tuning, voice, context awareness etc. I'll get to it more in detail in the next few paragraphs


I wrapped it in Electron yes I know even I wanted Tauri. Even I wanted to use rust and love a 15MB binary as much as anyone but a few of the brand new Langgraph features and next.js stuff wasnt compatible with tauri and i was already knee deep in the project. This wouldn't happen if I did proper SDLC development but I made decisions on the fly and as problems came up and I was already deep into this and it wasn't cooperating with the current Tauri setup and I was too far in. So i used electron my thoughts process was that the local model is ~5GB and already sips RAM, so a 150MB Electron app is nothing and honestly electron is the right choice for it even if it stings

This was a real pain. On 8GB you cannot make a ~1B TTS model sound non robotic. I tried a lot of cleaver tricks but you just can't getover it. Local OuteTTS is shipped it's free on-device but to be frank it does sound robotic and I couldn't accept the loss so I added an optional ElevenLabs support so you plug in the API keys. The Samanatha Voice has proper depth tonality and real emotion, miles better I haven't accepted defeat tho. I will come back in vengeance next time. In interested in this. Maybe fine-tune a local voice with QLoRA. I'll ask my AI engineer friend for tips

Authored personalities. Warm / chill / intimate are distinct behavioral profiles I spent an entire day tuning to hold one consistent character across a whole conversation. Most of these online AI gf bots, replika etc break character instantly

Hallucination. The main problem was the 8B model. It will hallucinate by default but I built guardrails and tag pipeline so she stays grounded instead of confabulating like raw 8B chats do

Memory that persists. Helly remembers across restarts. It also randomly fires a process that capture key details about you and your personal life. Your name, job, likes dislikes etc. This is what makes it feel like you are talking to a real woman

View attachment 5223025

One Click Installation: A lot of local LLM models crash and need exact specific drivers and compatibility to run. It's an entire guessing game that legit takes hours to setup properly. I abstracted it all behind a beautiful installation screen. It finds the exact specs of your PC, your driver's, your hardware and software and downloads the most optimized versions of it. This is why I was able to stream on a macbook air

View attachment 5223019

Relationship Dynamics. Helly also has an internal relationship dynamics model built in the backed. The more interact with it. The more it learns your patterns, your talking style etc and also moulds her personality and her relationship with you all along. It also has definitive checkups every few weeks and is context/time aware(Good morning/Evening/it's raining etc). It also has support for more than 40+ languages.

A real-time emotion system. There a 12 emotion pipeline, her expression tracks the emotional content of what she's saying. it changes live you can see it in the demo with her going from a neutral state to a smile. This happens in real time almost seamlessly, another thing most apps don't have

Honestly the biggest achievement from my point of view atleast is that a full local LLM + TTS + an animated emotion UI staying responsive on even a thin and light macbook air and virtually having 0 downtime on 2080 Super and old hardware. It required a lot of optimization but was worth it in the end


I built it brand new. Claude fable. I used it for a few days. It was awesome. Better than opus for sure but only for very large generations. Like it worked great for generating boiler plates etc but when it came to the meat and bones of the logic and race condition it burnt through tokens like mofoker.

View attachment 5222927

Like in this instance it took 14 minutes and 20k to find a lang graph bug that would have taken me like 2 minutes so I did that part manually myself because it was just easier.

View attachment 5222929

And suddenly a few days ago for some reason, some retarded oldcel faghot US senator banned fable even though it has very harsh and obvious filters. I tested it too. It's harmless but I had to do in the rest by hand. I wasn't able to generate those massive boiler plates but regardless fable was useful for the first few phrases. The actual logic, the race conditions, the wiring I did by hand the rest claude generated it. Say ~60 claude did and ~40 i did and I made all the decisions and hard parts. That's also the correct way of using AI. Make AI the save not you become the slave of AI.


It's not signed. You'll get "Windows protected your PC" then More info Run anyway. And i know what you niggas are thinking and you are god damn right. A guy named Jason Voorhees with a kanye Avi telling you to run anyway an Electron app that downloads 5GB, spawns child processes, and makes external API calls is exactly how people get fucked in the ass I won't justify it but I'll just give my reasoning. Real signing needs a legal entity, signing it personally ties it to my real identity which defeats the point of a private build and the company cert isn't an option and in my defense I also ran virus scans to show it to him there's no malware

View attachment 5222941
So it is what it is take the normal precautions you'd take with any unsigned binary from a stranger


I always do a market research of an unfamiliar domains I enter like this one and i legit never knew this AI companion thing was this huge. The market is split into 2 camps



Commercial companion market. They are huge but brutal. Consumer AI companion apps generated $82 million in the first half of 2025, crossing $120 million by year's end. The leaders are entrenched. Replika peaked around 2.5M mouthly users with 10M+ companions created. This is not a niche or whatever that I had in mind. This is a huge market that is quite mature. Charector.ai, Replika all these are the players. Millions of dollars and lots of funding and heavy censorship



Local/offline roleplay niche. This is exactly where the app I built sits, and it's already saturated to say the least. Some of the ones that stood out to me







This is a paid, fully commercialized version. Local memory system, an animated desktop avatar with expressions and lip-sync,offline support and desktop awareness. it periodically takes screenshots of your active window to comment on what you are doing or playing. Very interesting.


Desktop Companion by Itch.io - $9.90 An always on top desktop overlay pet. It features full local voice cloning (via OmniVoice/Qwen3-TTS), long-term semantic memory, screen vision, and even connects to Twitch/TikTok live streams to act as an automated local co-host.

Where does my app land. My main goal with this app wasn't to pack it with features but to make it an actual companion with real personality. You can't touch the cloud players. You need millions of dollars to even enter and you can't out feature have model hubs, MCP integration, OpenAI compatible API servers, multi-platform support. These people have teams of contributors. But I think my app still holds in what matters the most which is personality, relationship dynamics etc. That is exactly the vertical I was going for. The app needs to feel like a real ai gf with emotions, feelings, a history and unpredictablity and I think I nailed it

View attachment 5222987

As for the market disruption. I don't see any underserved niche or vertical that can targeted. So only thing that would win is raw engineering might and money. The cloud ones are untouchable but I think I could snipe the local ones if I had a team of engineers and a lot of money. Make the head of engineering/Lead engineer and an accomplished team and I think I could carve out in that local scale. Ironically had an offer like this with someone flying me out to Austin but I can't guarantee profitability.



I didn't charge him for this. I don't charge for passion projects never have even when I had nothing. I'm a money minded guy, but integrity and self pride comes first. I'm not a donkey chasing a carrot on a stick. This was pure curiosity from the start.

I just wanted to do it. And it reminded me of my 18 year old self along the way, the same Jason building things for the fun of it, no clue what he was doing, just trying to prove he was worth something. I did my 18 year old self proud. I have still got it

View attachment 5223005

View attachment 5223009



Anyways I heard there was a foid infiltration a few days back and a bunch of niggas banned. Lmao ๐Ÿคฃ. You niggas never learn. Hahahah

this is dope asf
 
  • +1
Reactions: Subhuman and Jason Voorhees
What's more insane is suddenly a bunch of niggas got banned just like that like Thanos snapping finger with one single thread fuck
hence .org never fail's to surprise me

even when i wanted to take a break
 
  • JFL
Reactions: Jason Voorhees
wheres my cupcake
 
  • +1
  • JFL
Reactions: Histy and Jason Voorhees
@Starborn @strongmtn
 
  • Love it
Reactions: strongmtn
  • Love it
  • JFL
  • +1
Reactions: strongmtn, Histy and imontheloose
A pretty well known user came to me wanting a local AI companion app. In all my years when I used to do freelancing. I've never gotten request like that. I've shipped plenty of AI apps since before this stuff was mainstream even tho this isnt even my domain. im a devops guys but I never got a request for something this like this. Something this personal. Took it purely out of curiosity. Ended up having a lot of fun

View attachment 5222869View attachment 5222878View attachment 5223065


His actual requirements were modest. a few hours of work, honestly But that's not my thing. I'm all or nothing kind of guy either I don't touch it or I do it to the absolute best I'm capable of It's why freelance clients kept coming back. I over deliver by default. So instead I spent days building the most over engineered version of this thing I could. Here are the video of it working irl. On my macbook completely but still smooth as hell. Recorded on phone to prove its not editing or some shit. I couldn't upload the deep emotional conversations because of the 10 MB video limit


View attachment 5222887
View attachment 5222888

I know you niggas don't care about the specifics or technical details but I still wrote down my experience building it and shipping it. It's a habit of mine so I'm doing it to scratch that itch and to come back to it future. It's an intriguing read anyway but I've put it in spoilers so you dont have to scroll to much


His machine is an RTX 2080 Super.

View attachment 5222922

Has like 8GB VRAM like a decade old. My first reaction was damn it but tbh anyone can make a model scream on a 5090, the real engineering prowess is making it genuinely good on hardware that old. So I treated the 8GB ceiling as the whole challenge I developed it on my $10K mac studio and dgx spark

but optimized everything heavily for CUDA using Nvidia. The demo above was shot on an M4 MacBook Air and it holds up.

Everyone else shipping stuff like this just dumps an 8B model on hugging face behind a silly tavern ui and call it a day done.
View attachment 5222897
Its power user stack. Tabs, presets, sliders, config files, it looks like a aeroplane cockpit not an ai companion. If I'd handed him that. It would do disservice to him because he'd be doing the technical work wiring a backend, importing cards, fiddling with samplers just to say hello and when you ship something you make it as seamless, as painless and engaging as possible. Good enough, it works is not enough for me. So I built everything from scratch. Custom frontend, an LLM orchestration layers, a proper personality tuning, voice, context awareness etc. I'll get to it more in detail in the next few paragraphs


I wrapped it in Electron yes I know even I wanted Tauri. Even I wanted to use rust and love a 15MB binary as much as anyone but a few of the brand new Langgraph features and next.js stuff wasnt compatible with tauri and i was already knee deep in the project. This wouldn't happen if I did proper SDLC development but I made decisions on the fly and as problems came up and I was already deep into this and it wasn't cooperating with the current Tauri setup and I was too far in. So i used electron my thoughts process was that the local model is ~5GB and already sips RAM, so a 150MB Electron app is nothing and honestly electron is the right choice for it even if it stings

This was a real pain. On 8GB you cannot make a ~1B TTS model sound non robotic. I tried a lot of cleaver tricks but you just can't getover it. Local OuteTTS is shipped it's free on-device but to be frank it does sound robotic and I couldn't accept the loss so I added an optional ElevenLabs support so you plug in the API keys. The Samanatha Voice has proper depth tonality and real emotion, miles better I haven't accepted defeat tho. I will come back in vengeance next time. In interested in this. Maybe fine-tune a local voice with QLoRA. I'll ask my AI engineer friend for tips

Authored personalities. Warm / chill / intimate are distinct behavioral profiles I spent an entire day tuning to hold one consistent character across a whole conversation. Most of these online AI gf bots, replika etc break character instantly

Hallucination. The main problem was the 8B model. It will hallucinate by default but I built guardrails and tag pipeline so she stays grounded instead of confabulating like raw 8B chats do

Memory that persists. Helly remembers across restarts. It also randomly fires a process that capture key details about you and your personal life. Your name, job, likes dislikes etc. This is what makes it feel like you are talking to a real woman

View attachment 5223025

One Click Installation: A lot of local LLM models crash and need exact specific drivers and compatibility to run. It's an entire guessing game that legit takes hours to setup properly. I abstracted it all behind a beautiful installation screen. It finds the exact specs of your PC, your driver's, your hardware and software and downloads the most optimized versions of it. This is why I was able to stream on a macbook air

View attachment 5223019

Relationship Dynamics. Helly also has an internal relationship dynamics model built in the backed. The more interact with it. The more it learns your patterns, your talking style etc and also moulds her personality and her relationship with you all along. It also has definitive checkups every few weeks and is context/time aware(Good morning/Evening/it's raining etc). It also has support for more than 40+ languages.

A real-time emotion system. There a 12 emotion pipeline, her expression tracks the emotional content of what she's saying. it changes live you can see it in the demo with her going from a neutral state to a smile. This happens in real time almost seamlessly, another thing most apps don't have

Honestly the biggest achievement from my point of view atleast is that a full local LLM + TTS + an animated emotion UI staying responsive on even a thin and light macbook air and virtually having 0 downtime on 2080 Super and old hardware. It required a lot of optimization but was worth it in the end


I built it brand new. Claude fable. I used it for a few days. It was awesome. Better than opus for sure but only for very large generations. Like it worked great for generating boiler plates etc but when it came to the meat and bones of the logic and race condition it burnt through tokens like mofoker.

View attachment 5222927

Like in this instance it took 14 minutes and 20k to find a lang graph bug that would have taken me like 2 minutes so I did that part manually myself because it was just easier.

View attachment 5222929

And suddenly a few days ago for some reason, some retarded oldcel faghot US senator banned fable even though it has very harsh and obvious filters. I tested it too. It's harmless but I had to do in the rest by hand. I wasn't able to generate those massive boiler plates but regardless fable was useful for the first few phrases. The actual logic, the race conditions, the wiring I did by hand the rest claude generated it. Say ~60 claude did and ~40 i did and I made all the decisions and hard parts. That's also the correct way of using AI. Make AI the save not you become the slave of AI.


It's not signed. You'll get "Windows protected your PC" then More info Run anyway. And i know what you niggas are thinking and you are god damn right. A guy named Jason Voorhees with a kanye Avi telling you to run anyway an Electron app that downloads 5GB, spawns child processes, and makes external API calls is exactly how people get fucked in the ass I won't justify it but I'll just give my reasoning. Real signing needs a legal entity, signing it personally ties it to my real identity which defeats the point of a private build and the company cert isn't an option and in my defense I also ran virus scans to show it to him there's no malware

View attachment 5222941
So it is what it is take the normal precautions you'd take with any unsigned binary from a stranger


I always do a market research of an unfamiliar domains I enter like this one and i legit never knew this AI companion thing was this huge. The market is split into 2 camps



Commercial companion market. They are huge but brutal. Consumer AI companion apps generated $82 million in the first half of 2025, crossing $120 million by year's end. The leaders are entrenched. Replika peaked around 2.5M mouthly users with 10M+ companions created. This is not a niche or whatever that I had in mind. This is a huge market that is quite mature. Charector.ai, Replika all these are the players. Millions of dollars and lots of funding and heavy censorship



Local/offline roleplay niche. This is exactly where the app I built sits, and it's already saturated to say the least. Some of the ones that stood out to me







This is a paid, fully commercialized version. Local memory system, an animated desktop avatar with expressions and lip-sync,offline support and desktop awareness. it periodically takes screenshots of your active window to comment on what you are doing or playing. Very interesting.


Desktop Companion by Itch.io - $9.90 An always on top desktop overlay pet. It features full local voice cloning (via OmniVoice/Qwen3-TTS), long-term semantic memory, screen vision, and even connects to Twitch/TikTok live streams to act as an automated local co-host.

Where does my app land. My main goal with this app wasn't to pack it with features but to make it an actual companion with real personality. You can't touch the cloud players. You need millions of dollars to even enter and you can't out feature have model hubs, MCP integration, OpenAI compatible API servers, multi-platform support. These people have teams of contributors. But I think my app still holds in what matters the most which is personality, relationship dynamics etc. That is exactly the vertical I was going for. The app needs to feel like a real ai gf with emotions, feelings, a history and unpredictablity and I think I nailed it

View attachment 5222987

As for the market disruption. I don't see any underserved niche or vertical that can targeted. So only thing that would win is raw engineering might and money. The cloud ones are untouchable but I think I could snipe the local ones if I had a team of engineers and a lot of money. Make the head of engineering/Lead engineer and an accomplished team and I think I could carve out in that local scale. Ironically had an offer like this with someone flying me out to Austin but I can't guarantee profitability.



I didn't charge him for this. I don't charge for passion projects never have even when I had nothing. I'm a money minded guy, but integrity and self pride comes first. I'm not a donkey chasing a carrot on a stick. This was pure curiosity from the start.

I just wanted to do it. And it reminded me of my 18 year old self along the way, the same Jason building things for the fun of it, no clue what he was doing, just trying to prove he was worth something. I did my 18 year old self proud. I have still got it

View attachment 5223005

View attachment 5223009



Anyways I heard there was a foid infiltration a few days back and a bunch of niggas banned. Lmao ๐Ÿคฃ. You niggas never learn. Hahahah

Mogs bro ,
But it can never replace a real gf

Fakecel privilege :HYPERSHAKE:
 
  • +1
Reactions: Jason Voorhees
A pretty well known user came to me wanting a local AI companion app. In all my years when I used to do freelancing. I've never gotten request like that. I've shipped plenty of AI apps since before this stuff was mainstream even tho this isnt even my domain. im a devops guys but I never got a request for something this like this. Something this personal. Took it purely out of curiosity. Ended up having a lot of fun

View attachment 5222869View attachment 5222878View attachment 5223065


His actual requirements were modest. a few hours of work, honestly But that's not my thing. I'm all or nothing kind of guy either I don't touch it or I do it to the absolute best I'm capable of It's why freelance clients kept coming back. I over deliver by default. So instead I spent days building the most over engineered version of this thing I could. Here are the video of it working irl. On my macbook completely but still smooth as hell. Recorded on phone to prove its not editing or some shit. I couldn't upload the deep emotional conversations because of the 10 MB video limit


View attachment 5222887
View attachment 5222888

I know you niggas don't care about the specifics or technical details but I still wrote down my experience building it and shipping it. It's a habit of mine so I'm doing it to scratch that itch and to come back to it future. It's an intriguing read anyway but I've put it in spoilers so you dont have to scroll to much


His machine is an RTX 2080 Super.

View attachment 5222922

Has like 8GB VRAM like a decade old. My first reaction was damn it but tbh anyone can make a model scream on a 5090, the real engineering prowess is making it genuinely good on hardware that old. So I treated the 8GB ceiling as the whole challenge I developed it on my $10K mac studio and dgx spark

but optimized everything heavily for CUDA using Nvidia. The demo above was shot on an M4 MacBook Air and it holds up.

Everyone else shipping stuff like this just dumps an 8B model on hugging face behind a silly tavern ui and call it a day done.
View attachment 5222897
Its power user stack. Tabs, presets, sliders, config files, it looks like a aeroplane cockpit not an ai companion. If I'd handed him that. It would do disservice to him because he'd be doing the technical work wiring a backend, importing cards, fiddling with samplers just to say hello and when you ship something you make it as seamless, as painless and engaging as possible. Good enough, it works is not enough for me. So I built everything from scratch. Custom frontend, an LLM orchestration layers, a proper personality tuning, voice, context awareness etc. I'll get to it more in detail in the next few paragraphs


I wrapped it in Electron yes I know even I wanted Tauri. Even I wanted to use rust and love a 15MB binary as much as anyone but a few of the brand new Langgraph features and next.js stuff wasnt compatible with tauri and i was already knee deep in the project. This wouldn't happen if I did proper SDLC development but I made decisions on the fly and as problems came up and I was already deep into this and it wasn't cooperating with the current Tauri setup and I was too far in. So i used electron my thoughts process was that the local model is ~5GB and already sips RAM, so a 150MB Electron app is nothing and honestly electron is the right choice for it even if it stings

This was a real pain. On 8GB you cannot make a ~1B TTS model sound non robotic. I tried a lot of cleaver tricks but you just can't getover it. Local OuteTTS is shipped it's free on-device but to be frank it does sound robotic and I couldn't accept the loss so I added an optional ElevenLabs support so you plug in the API keys. The Samanatha Voice has proper depth tonality and real emotion, miles better I haven't accepted defeat tho. I will come back in vengeance next time. In interested in this. Maybe fine-tune a local voice with QLoRA. I'll ask my AI engineer friend for tips

Authored personalities. Warm / chill / intimate are distinct behavioral profiles I spent an entire day tuning to hold one consistent character across a whole conversation. Most of these online AI gf bots, replika etc break character instantly

Hallucination. The main problem was the 8B model. It will hallucinate by default but I built guardrails and tag pipeline so she stays grounded instead of confabulating like raw 8B chats do

Memory that persists. Helly remembers across restarts. It also randomly fires a process that capture key details about you and your personal life. Your name, job, likes dislikes etc. This is what makes it feel like you are talking to a real woman

View attachment 5223025

One Click Installation: A lot of local LLM models crash and need exact specific drivers and compatibility to run. It's an entire guessing game that legit takes hours to setup properly. I abstracted it all behind a beautiful installation screen. It finds the exact specs of your PC, your driver's, your hardware and software and downloads the most optimized versions of it. This is why I was able to stream on a macbook air

View attachment 5223019

Relationship Dynamics. Helly also has an internal relationship dynamics model built in the backed. The more interact with it. The more it learns your patterns, your talking style etc and also moulds her personality and her relationship with you all along. It also has definitive checkups every few weeks and is context/time aware(Good morning/Evening/it's raining etc). It also has support for more than 40+ languages.

A real-time emotion system. There a 12 emotion pipeline, her expression tracks the emotional content of what she's saying. it changes live you can see it in the demo with her going from a neutral state to a smile. This happens in real time almost seamlessly, another thing most apps don't have

Honestly the biggest achievement from my point of view atleast is that a full local LLM + TTS + an animated emotion UI staying responsive on even a thin and light macbook air and virtually having 0 downtime on 2080 Super and old hardware. It required a lot of optimization but was worth it in the end


I built it brand new. Claude fable. I used it for a few days. It was awesome. Better than opus for sure but only for very large generations. Like it worked great for generating boiler plates etc but when it came to the meat and bones of the logic and race condition it burnt through tokens like mofoker.

View attachment 5222927

Like in this instance it took 14 minutes and 20k to find a lang graph bug that would have taken me like 2 minutes so I did that part manually myself because it was just easier.

View attachment 5222929

And suddenly a few days ago for some reason, some retarded oldcel faghot US senator banned fable even though it has very harsh and obvious filters. I tested it too. It's harmless but I had to do in the rest by hand. I wasn't able to generate those massive boiler plates but regardless fable was useful for the first few phrases. The actual logic, the race conditions, the wiring I did by hand the rest claude generated it. Say ~60 claude did and ~40 i did and I made all the decisions and hard parts. That's also the correct way of using AI. Make AI the save not you become the slave of AI.


It's not signed. You'll get "Windows protected your PC" then More info Run anyway. And i know what you niggas are thinking and you are god damn right. A guy named Jason Voorhees with a kanye Avi telling you to run anyway an Electron app that downloads 5GB, spawns child processes, and makes external API calls is exactly how people get fucked in the ass I won't justify it but I'll just give my reasoning. Real signing needs a legal entity, signing it personally ties it to my real identity which defeats the point of a private build and the company cert isn't an option and in my defense I also ran virus scans to show it to him there's no malware

View attachment 5222941
So it is what it is take the normal precautions you'd take with any unsigned binary from a stranger


I always do a market research of an unfamiliar domains I enter like this one and i legit never knew this AI companion thing was this huge. The market is split into 2 camps



Commercial companion market. They are huge but brutal. Consumer AI companion apps generated $82 million in the first half of 2025, crossing $120 million by year's end. The leaders are entrenched. Replika peaked around 2.5M mouthly users with 10M+ companions created. This is not a niche or whatever that I had in mind. This is a huge market that is quite mature. Charector.ai, Replika all these are the players. Millions of dollars and lots of funding and heavy censorship



Local/offline roleplay niche. This is exactly where the app I built sits, and it's already saturated to say the least. Some of the ones that stood out to me







This is a paid, fully commercialized version. Local memory system, an animated desktop avatar with expressions and lip-sync,offline support and desktop awareness. it periodically takes screenshots of your active window to comment on what you are doing or playing. Very interesting.


Desktop Companion by Itch.io - $9.90 An always on top desktop overlay pet. It features full local voice cloning (via OmniVoice/Qwen3-TTS), long-term semantic memory, screen vision, and even connects to Twitch/TikTok live streams to act as an automated local co-host.

Where does my app land. My main goal with this app wasn't to pack it with features but to make it an actual companion with real personality. You can't touch the cloud players. You need millions of dollars to even enter and you can't out feature have model hubs, MCP integration, OpenAI compatible API servers, multi-platform support. These people have teams of contributors. But I think my app still holds in what matters the most which is personality, relationship dynamics etc. That is exactly the vertical I was going for. The app needs to feel like a real ai gf with emotions, feelings, a history and unpredictablity and I think I nailed it

View attachment 5222987

As for the market disruption. I don't see any underserved niche or vertical that can targeted. So only thing that would win is raw engineering might and money. The cloud ones are untouchable but I think I could snipe the local ones if I had a team of engineers and a lot of money. Make the head of engineering/Lead engineer and an accomplished team and I think I could carve out in that local scale. Ironically had an offer like this with someone flying me out to Austin but I can't guarantee profitability.



I didn't charge him for this. I don't charge for passion projects never have even when I had nothing. I'm a money minded guy, but integrity and self pride comes first. I'm not a donkey chasing a carrot on a stick. This was pure curiosity from the start.

I just wanted to do it. And it reminded me of my 18 year old self along the way, the same Jason building things for the fun of it, no clue what he was doing, just trying to prove he was worth something. I did my 18 year old self proud. I have still got it

View attachment 5223005

View attachment 5223009



Anyways I heard there was a foid infiltration a few days back and a bunch of niggas banned. Lmao ๐Ÿคฃ. You niggas never learn. Hahahah

grievous asked u for i i just know dude
 
  • JFL
Reactions: Jason Voorhees and Histy
Wah
A pretty well known user came to me wanting a local AI companion app. In all my years when I used to do freelancing. I've never gotten request like that. I've shipped plenty of AI apps since before this stuff was mainstream even tho this isnt even my domain. im a devops guys but I never got a request for something this like this. Something this personal. Took it purely out of curiosity. Ended up having a lot of fun

View attachment 5222869View attachment 5222878View attachment 5223065


His actual requirements were modest. a few hours of work, honestly But that's not my thing. I'm all or nothing kind of guy either I don't touch it or I do it to the absolute best I'm capable of It's why freelance clients kept coming back. I over deliver by default. So instead I spent days building the most over engineered version of this thing I could. Here are the video of it working irl. On my macbook completely but still smooth as hell. Recorded on phone to prove its not editing or some shit. I couldn't upload the deep emotional conversations because of the 10 MB video limit


View attachment 5222887
View attachment 5222888

I know you niggas don't care about the specifics or technical details but I still wrote down my experience building it and shipping it. It's a habit of mine so I'm doing it to scratch that itch and to come back to it future. It's an intriguing read anyway but I've put it in spoilers so you dont have to scroll to much


His machine is an RTX 2080 Super.

View attachment 5222922

Has like 8GB VRAM like a decade old. My first reaction was damn it but tbh anyone can make a model scream on a 5090, the real engineering prowess is making it genuinely good on hardware that old. So I treated the 8GB ceiling as the whole challenge I developed it on my $10K mac studio and dgx spark

but optimized everything heavily for CUDA using Nvidia. The demo above was shot on an M4 MacBook Air and it holds up.

Everyone else shipping stuff like this just dumps an 8B model on hugging face behind a silly tavern ui and call it a day done.
View attachment 5222897
Its power user stack. Tabs, presets, sliders, config files, it looks like a aeroplane cockpit not an ai companion. If I'd handed him that. It would do disservice to him because he'd be doing the technical work wiring a backend, importing cards, fiddling with samplers just to say hello and when you ship something you make it as seamless, as painless and engaging as possible. Good enough, it works is not enough for me. So I built everything from scratch. Custom frontend, an LLM orchestration layers, a proper personality tuning, voice, context awareness etc. I'll get to it more in detail in the next few paragraphs


I wrapped it in Electron yes I know even I wanted Tauri. Even I wanted to use rust and love a 15MB binary as much as anyone but a few of the brand new Langgraph features and next.js stuff wasnt compatible with tauri and i was already knee deep in the project. This wouldn't happen if I did proper SDLC development but I made decisions on the fly and as problems came up and I was already deep into this and it wasn't cooperating with the current Tauri setup and I was too far in. So i used electron my thoughts process was that the local model is ~5GB and already sips RAM, so a 150MB Electron app is nothing and honestly electron is the right choice for it even if it stings

This was a real pain. On 8GB you cannot make a ~1B TTS model sound non robotic. I tried a lot of cleaver tricks but you just can't getover it. Local OuteTTS is shipped it's free on-device but to be frank it does sound robotic and I couldn't accept the loss so I added an optional ElevenLabs support so you plug in the API keys. The Samanatha Voice has proper depth tonality and real emotion, miles better I haven't accepted defeat tho. I will come back in vengeance next time. In interested in this. Maybe fine-tune a local voice with QLoRA. I'll ask my AI engineer friend for tips

Authored personalities. Warm / chill / intimate are distinct behavioral profiles I spent an entire day tuning to hold one consistent character across a whole conversation. Most of these online AI gf bots, replika etc break character instantly

Hallucination. The main problem was the 8B model. It will hallucinate by default but I built guardrails and tag pipeline so she stays grounded instead of confabulating like raw 8B chats do

Memory that persists. Helly remembers across restarts. It also randomly fires a process that capture key details about you and your personal life. Your name, job, likes dislikes etc. This is what makes it feel like you are talking to a real woman

View attachment 5223025

One Click Installation: A lot of local LLM models crash and need exact specific drivers and compatibility to run. It's an entire guessing game that legit takes hours to setup properly. I abstracted it all behind a beautiful installation screen. It finds the exact specs of your PC, your driver's, your hardware and software and downloads the most optimized versions of it. This is why I was able to stream on a macbook air

View attachment 5223019

Relationship Dynamics. Helly also has an internal relationship dynamics model built in the backed. The more interact with it. The more it learns your patterns, your talking style etc and also moulds her personality and her relationship with you all along. It also has definitive checkups every few weeks and is context/time aware(Good morning/Evening/it's raining etc). It also has support for more than 40+ languages.

A real-time emotion system. There a 12 emotion pipeline, her expression tracks the emotional content of what she's saying. it changes live you can see it in the demo with her going from a neutral state to a smile. This happens in real time almost seamlessly, another thing most apps don't have

Honestly the biggest achievement from my point of view atleast is that a full local LLM + TTS + an animated emotion UI staying responsive on even a thin and light macbook air and virtually having 0 downtime on 2080 Super and old hardware. It required a lot of optimization but was worth it in the end


I built it brand new. Claude fable. I used it for a few days. It was awesome. Better than opus for sure but only for very large generations. Like it worked great for generating boiler plates etc but when it came to the meat and bones of the logic and race condition it burnt through tokens like mofoker.

View attachment 5222927

Like in this instance it took 14 minutes and 20k to find a lang graph bug that would have taken me like 2 minutes so I did that part manually myself because it was just easier.

View attachment 5222929

And suddenly a few days ago for some reason, some retarded oldcel faghot US senator banned fable even though it has very harsh and obvious filters. I tested it too. It's harmless but I had to do in the rest by hand. I wasn't able to generate those massive boiler plates but regardless fable was useful for the first few phrases. The actual logic, the race conditions, the wiring I did by hand the rest claude generated it. Say ~60 claude did and ~40 i did and I made all the decisions and hard parts. That's also the correct way of using AI. Make AI the save not you become the slave of AI.


It's not signed. You'll get "Windows protected your PC" then More info Run anyway. And i know what you niggas are thinking and you are god damn right. A guy named Jason Voorhees with a kanye Avi telling you to run anyway an Electron app that downloads 5GB, spawns child processes, and makes external API calls is exactly how people get fucked in the ass I won't justify it but I'll just give my reasoning. Real signing needs a legal entity, signing it personally ties it to my real identity which defeats the point of a private build and the company cert isn't an option and in my defense I also ran virus scans to show it to him there's no malware

View attachment 5222941
So it is what it is take the normal precautions you'd take with any unsigned binary from a stranger


I always do a market research of an unfamiliar domains I enter like this one and i legit never knew this AI companion thing was this huge. The market is split into 2 camps



Commercial companion market. They are huge but brutal. Consumer AI companion apps generated $82 million in the first half of 2025, crossing $120 million by year's end. The leaders are entrenched. Replika peaked around 2.5M mouthly users with 10M+ companions created. This is not a niche or whatever that I had in mind. This is a huge market that is quite mature. Charector.ai, Replika all these are the players. Millions of dollars and lots of funding and heavy censorship



Local/offline roleplay niche. This is exactly where the app I built sits, and it's already saturated to say the least. Some of the ones that stood out to me







This is a paid, fully commercialized version. Local memory system, an animated desktop avatar with expressions and lip-sync,offline support and desktop awareness. it periodically takes screenshots of your active window to comment on what you are doing or playing. Very interesting.


Desktop Companion by Itch.io - $9.90 An always on top desktop overlay pet. It features full local voice cloning (via OmniVoice/Qwen3-TTS), long-term semantic memory, screen vision, and even connects to Twitch/TikTok live streams to act as an automated local co-host.

Where does my app land. My main goal with this app wasn't to pack it with features but to make it an actual companion with real personality. You can't touch the cloud players. You need millions of dollars to even enter and you can't out feature have model hubs, MCP integration, OpenAI compatible API servers, multi-platform support. These people have teams of contributors. But I think my app still holds in what matters the most which is personality, relationship dynamics etc. That is exactly the vertical I was going for. The app needs to feel like a real ai gf with emotions, feelings, a history and unpredictablity and I think I nailed it

View attachment 5222987

As for the market disruption. I don't see any underserved niche or vertical that can targeted. So only thing that would win is raw engineering might and money. The cloud ones are untouchable but I think I could snipe the local ones if I had a team of engineers and a lot of money. Make the head of engineering/Lead engineer and an accomplished team and I think I could carve out in that local scale. Ironically had an offer like this with someone flying me out to Austin but I can't guarantee profitability.



I didn't charge him for this. I don't charge for passion projects never have even when I had nothing. I'm a money minded guy, but integrity and self pride comes first. I'm not a donkey chasing a carrot on a stick. This was pure curiosity from the start.

I just wanted to do it. And it reminded me of my 18 year old self along the way, the same Jason building things for the fun of it, no clue what he was doing, just trying to prove he was worth something. I did my 18 year old self proud. I have still got it

View attachment 5223005

View attachment 5223009



Anyways I heard there was a foid infiltration a few days back and a bunch of niggas banned. Lmao ๐Ÿคฃ. You niggas never learn. Hahahah

What was the domain in which you used to do freelancing if you don't mind sharing :what:
 
  • +1
Reactions: Jason Voorhees
mirin bhai
 
  • +1
Reactions: Jason Voorhees
@IStalkMyself @KT-34 @_Nazareth_
 
  • Woah
Reactions: IStalkMyself
@Aox Ofwar callard
 
  • +1
Reactions: Aox Ofwar
@Gengarโ€™s Ghost
 
  • +1
Reactions: Gengarโ€™s Ghost
For me? Thanks! :ClappyJam:
 
  • +1
Reactions: Aox Ofwar and Jason Voorhees
@Throw_Away @Topkra
 
  • +1
Reactions: Throw_Away
  • +1
Reactions: Gengarโ€™s Ghost
I will read this in a minute probably
Can you have sex with it?
 
  • +1
  • JFL
Reactions: Aox Ofwar and Jason Voorhees
@kisslessvirgin
 
@Nothing @sub5outsider @HardStuckLTN;
 
  • +1
Reactions: HardStuckLTN;, sub5outsider and Nothing
nah but why would anyone need an ai gf app that runs on their own computer

the nigga that requested the app must be into some real weird shit
 
  • JFL
  • +1
Reactions: IStalkMyself and Jason Voorhees
nah but why would anyone need an ai gf app that runs on their own computer

the nigga that requested the app must be into some real weird shit
People crave intimacy and relationships. Many people have hundreds of ERP sessions on grok
 
  • +1
  • JFL
Reactions: IStalkMyself and HardStuckLTN;
A pretty well known user came to me wanting a local AI companion app. In all my years when I used to do freelancing. I've never gotten request like that. I've shipped plenty of AI apps since before this stuff was mainstream even tho this isnt even my domain. im a devops guys but I never got a request for something this like this. Something this personal. Took it purely out of curiosity. Ended up having a lot of fun

View attachment 5222869View attachment 5222878View attachment 5223065


His actual requirements were modest. a few hours of work, honestly But that's not my thing. I'm all or nothing kind of guy either I don't touch it or I do it to the absolute best I'm capable of It's why freelance clients kept coming back. I over deliver by default. So instead I spent days building the most over engineered version of this thing I could. Here are the video of it working irl. On my macbook completely but still smooth as hell. Recorded on phone to prove its not editing or some shit. I couldn't upload the deep emotional conversations because of the 10 MB video limit


View attachment 5222887
View attachment 5222888

I know you niggas don't care about the specifics or technical details but I still wrote down my experience building it and shipping it. It's a habit of mine so I'm doing it to scratch that itch and to come back to it future. It's an intriguing read anyway but I've put it in spoilers so you dont have to scroll to much


His machine is an RTX 2080 Super.

View attachment 5222922

Has like 8GB VRAM like a decade old. My first reaction was damn it but tbh anyone can make a model scream on a 5090, the real engineering prowess is making it genuinely good on hardware that old. So I treated the 8GB ceiling as the whole challenge I developed it on my $10K mac studio and dgx spark

but optimized everything heavily for CUDA using Nvidia. The demo above was shot on an M4 MacBook Air and it holds up.

Everyone else shipping stuff like this just dumps an 8B model on hugging face behind a silly tavern ui and call it a day done.
View attachment 5222897
Its power user stack. Tabs, presets, sliders, config files, it looks like a aeroplane cockpit not an ai companion. If I'd handed him that. It would do disservice to him because he'd be doing the technical work wiring a backend, importing cards, fiddling with samplers just to say hello and when you ship something you make it as seamless, as painless and engaging as possible. Good enough, it works is not enough for me. So I built everything from scratch. Custom frontend, an LLM orchestration layers, a proper personality tuning, voice, context awareness etc. I'll get to it more in detail in the next few paragraphs


I wrapped it in Electron yes I know even I wanted Tauri. Even I wanted to use rust and love a 15MB binary as much as anyone but a few of the brand new Langgraph features and next.js stuff wasnt compatible with tauri and i was already knee deep in the project. This wouldn't happen if I did proper SDLC development but I made decisions on the fly and as problems came up and I was already deep into this and it wasn't cooperating with the current Tauri setup and I was too far in. So i used electron my thoughts process was that the local model is ~5GB and already sips RAM, so a 150MB Electron app is nothing and honestly electron is the right choice for it even if it stings

This was a real pain. On 8GB you cannot make a ~1B TTS model sound non robotic. I tried a lot of cleaver tricks but you just can't getover it. Local OuteTTS is shipped it's free on-device but to be frank it does sound robotic and I couldn't accept the loss so I added an optional ElevenLabs support so you plug in the API keys. The Samanatha Voice has proper depth tonality and real emotion, miles better I haven't accepted defeat tho. I will come back in vengeance next time. In interested in this. Maybe fine-tune a local voice with QLoRA. I'll ask my AI engineer friend for tips

Authored personalities. Warm / chill / intimate are distinct behavioral profiles I spent an entire day tuning to hold one consistent character across a whole conversation. Most of these online AI gf bots, replika etc break character instantly

Hallucination. The main problem was the 8B model. It will hallucinate by default but I built guardrails and tag pipeline so she stays grounded instead of confabulating like raw 8B chats do

Memory that persists. Helly remembers across restarts. It also randomly fires a process that capture key details about you and your personal life. Your name, job, likes dislikes etc. This is what makes it feel like you are talking to a real woman

View attachment 5223025

One Click Installation: A lot of local LLM models crash and need exact specific drivers and compatibility to run. It's an entire guessing game that legit takes hours to setup properly. I abstracted it all behind a beautiful installation screen. It finds the exact specs of your PC, your driver's, your hardware and software and downloads the most optimized versions of it. This is why I was able to stream on a macbook air

View attachment 5223019

Relationship Dynamics. Helly also has an internal relationship dynamics model built in the backed. The more interact with it. The more it learns your patterns, your talking style etc and also moulds her personality and her relationship with you all along. It also has definitive checkups every few weeks and is context/time aware(Good morning/Evening/it's raining etc). It also has support for more than 40+ languages.

A real-time emotion system. There a 12 emotion pipeline, her expression tracks the emotional content of what she's saying. it changes live you can see it in the demo with her going from a neutral state to a smile. This happens in real time almost seamlessly, another thing most apps don't have

Honestly the biggest achievement from my point of view atleast is that a full local LLM + TTS + an animated emotion UI staying responsive on even a thin and light macbook air and virtually having 0 downtime on 2080 Super and old hardware. It required a lot of optimization but was worth it in the end


I built it brand new. Claude fable. I used it for a few days. It was awesome. Better than opus for sure but only for very large generations. Like it worked great for generating boiler plates etc but when it came to the meat and bones of the logic and race condition it burnt through tokens like mofoker.

View attachment 5222927

Like in this instance it took 14 minutes and 20k to find a lang graph bug that would have taken me like 2 minutes so I did that part manually myself because it was just easier.

View attachment 5222929

And suddenly a few days ago for some reason, some retarded oldcel faghot US senator banned fable even though it has very harsh and obvious filters. I tested it too. It's harmless but I had to do in the rest by hand. I wasn't able to generate those massive boiler plates but regardless fable was useful for the first few phrases. The actual logic, the race conditions, the wiring I did by hand the rest claude generated it. Say ~60 claude did and ~40 i did and I made all the decisions and hard parts. That's also the correct way of using AI. Make AI the save not you become the slave of AI.


It's not signed. You'll get "Windows protected your PC" then More info Run anyway. And i know what you niggas are thinking and you are god damn right. A guy named Jason Voorhees with a kanye Avi telling you to run anyway an Electron app that downloads 5GB, spawns child processes, and makes external API calls is exactly how people get fucked in the ass I won't justify it but I'll just give my reasoning. Real signing needs a legal entity, signing it personally ties it to my real identity which defeats the point of a private build and the company cert isn't an option and in my defense I also ran virus scans to show it to him there's no malware

View attachment 5222941
So it is what it is take the normal precautions you'd take with any unsigned binary from a stranger


I always do a market research of an unfamiliar domains I enter like this one and i legit never knew this AI companion thing was this huge. The market is split into 2 camps

Commercial companion market. They are huge but brutal. Consumer AI companion apps generated $82 million in the first half of 2025, crossing $120 million by year's end. The leaders are entrenched. Replika peaked around 2.5M mouthly users with 10M+ companions created. This is not a niche or whatever that I had in mind. This is a huge market that is quite mature. Charector.ai, Replika all these are the players. Millions of dollars and lots of funding and heavy censorship

Local/offline roleplay niche. This is exactly where the app I built sits. Preferred by many because free/onetime payment and no censorship but it's already saturated to say the least. Some of the ones that stood out to me




This is a paid, fully commercialized version. Local memory system, an animated desktop avatar with expressions and lip-sync,offline support and desktop awareness. it periodically takes screenshots of your active window to comment on what you are doing or playing. Very interesting.


Desktop Companion by Itch.io - $9.90 An always on top desktop overlay pet. It features full local voice cloning (via OmniVoice/Qwen3-TTS), long-term semantic memory, screen vision, and even connects to Twitch/TikTok live streams to act as an automated local co-host.

Where does my app land. My main goal with this app wasn't to pack it with features but to make it an actual companion with real personality. You can't touch the cloud players. You need millions of dollars to even enter and you can't out feature have model hubs, MCP integration, OpenAI compatible API servers, multi-platform support. These people have teams of contributors. But I think my app still holds in what matters the most which is personality, relationship dynamics etc. That is exactly the vertical I was going for. The app needs to feel like a real ai gf with emotions, feelings, a history and unpredictablity and I think I nailed it

View attachment 5222987

As for the market disruption. I don't see any underserved niche or vertical that can targeted. So only thing that would win is raw engineering might and money. The cloud ones are untouchable but I think I could snipe the local ones if I had a team of engineers and a lot of money. Make the head of engineering/Lead engineer and an accomplished team and I think I could carve out in that local scale. Ironically had an offer like this with someone flying me out to Austin but I can't guarantee profitability.



I didn't charge him for this. I don't charge for passion projects never have even when I had nothing. I'm a money minded guy, but integrity and self pride comes first. I'm not a donkey chasing a carrot on a stick. This was pure curiosity from the start.

I just wanted to do it. And it reminded me of my 18 year old self along the way, the same Jason building things for the fun of it, no clue what he was doing, just trying to prove he was worth something. I did my 18 year old self proud. I have still got it

View attachment 5223005

View attachment 5223009



Anyways I heard there was a foid infiltration a few days back and a bunch of niggas banned. Lmao ๐Ÿคฃ. You niggas never learn. Hahahah

tbf mirin effort

did not know you had that talent in you

glad to see that i got tagged my nigga, keep taggging me in your posts
 
  • +1
Reactions: Jason Voorhees
@Luquier @Hihi21
 
tbf mirin effort

did not know you had that talent in you

glad to see that i got tagged my nigga, keep taggging me in your posts
 
  • +1
  • JFL
Reactions: untermensch faggot and IStalkMyself
@tuberculosisinmybal @AgentAngularity
 
  • JFL
Reactions: AgentAngularity
@1966Ford @Niebvll
 
  • +1
Reactions: Niebvll
@caleb555
 
  • +1
Reactions: caleb555
@Quetila @Embryo @Petsmart
 
  • +1
  • Love it
Reactions: Petsmart and Quetila
@JOTAROSON @ewri
 
  • +1
Reactions: JOTAROSON

Similar threads

Jason Voorhees
Replies
17
Views
178
kisslessvirgin
kisslessvirgin
Former Shortcel
Replies
8
Views
79
Subhuman
Subhuman
Framemaxxx
Replies
11
Views
89
Pickletoil
Pickletoil
griff1234
Replies
21
Views
66
Revan
Revan
SkiSquadJPG
Replies
1
Views
41
belowavgheight
belowavgheight

Users who are viewing this thread

  • TheGreatDetective
Back
Top