Mattgsa 315 gallon build

[That Crusso Kid]

Over the past few weeks, I built my own AI assistant called Jarvise. It runs entirely on hardware I already own. No cloud, no subscriptions, no sending data off to someone else’s servers. It’s basically my own private version of ChatGPT, except it lives in my closet.

It runs on a pair of high end GPUs I had sitting around, and together they power a language model that’s honestly on par with the stuff people pay for. When I ask it something, it doesn’t just respond, it actually thinks through the problem and takes action.

That’s the part that changes everything. Most AI tools tell you what to do. Jarvise just does it.

If I ask it to build a web app, it writes the code, tests it, and deploys it on my network without me touching anything. I already put that to the test. It built a fully working task manager from scratch and shipped it on its own.

It can also handle images. Drop in a screenshot, photo, or diagram, and it switches to a vision model, figures it out, and switches back in a couple seconds. It feels seamless.

On top of that, it has access to my other machines. It knows all five computers on my network and can log into them, run commands, move files, restart services, basically act like an automated sysadmin while it’s solving whatever I asked.

Everything streams back in real time through a clean chat interface. It supports file uploads, formatted output, images, and it keeps track of my setup so it always has context.

Short version: I built a fully private, self hosted AI that runs on my own hardware, understands my environment, writes and deploys real software, and never sends my data anywhere else.

Holy cow, that sounds truly amazing! Did you write all the software/code for it? I am guessing you must have if it lives in it's own environment.

And now the big question: can it keep track of your kids? lol

 

[mattgsa]

Holy cow, that sounds truly amazing! Did you write all the software/code for it? I am guessing you must have if it lives in it's own environment.

And now the big question: can it keep track of your kids? lol

It can't track the kids yet, but that is in the plans, but it does try to write code, its a work in progress

 

[That Crusso Kid]

It can't track the kids yet, but that is in the plans, but it does try to write code, its a work in progress

Wow, man, just WOW!

I cannot help but wonder which GPU's you're using. Outside of that, I am pretty clueless about how one would go about setting something like what you have done up.

 

[mattgsa]

Wow, man, just WOW!

I cannot help but wonder which GPU's you're using. Outside of that, I am pretty clueless about how one would go about setting something like what you have done up.

I'm running dual 5080's

 

[mattgsa]

Right now I'm stuck here, the vision model trys to send the results to the LLM before its loaded and thus errors out

 

[That Crusso Kid]

I'm running dual 5080's

And you just happened to have a couple of them laying around? Can you please send me whatever other stuff you just happen to have laying around?

 

[mattgsa]

Right now I'm stuck here, the vision model trys to send the results to the LLM before its loaded and thus errors out

fixed it

 

[That Crusso Kid]

Dude... you are the MAN!

(Whatever you do, don't feed a pic of me into that thing. You'll break it!)

 

[mattgsa]

And you just happened to have a couple of them laying around? Can you please send me whatever other stuff you just happen to have laying around?

I actually built the whole setup specifically for this. I started with a single GPU, an AMD Ryzen 9, and 64GB of RAM. It worked, but it was pretty limited and struggled to keep up with what I wanted to do.

So I scaled it up. I added another 128GB of RAM and a second GPU, which made a huge difference in performance.

Getting there wasn’t exactly straightforward though. I went through four different motherboards before finding one that could properly support both GPUs. On top of that, I had to upgrade to a 1600 watt power supply and move everything into a much larger case to handle the power and cooling. I now have my eye on the Nvidia 6000 pro's but thats a little bit of an investment, but I would love to see what two of those would do for my ai.

Now it’s finally where I wanted it to be.

 

[mattgsa]

Gave it homework since I'm going to bed, LOL

 

[That Crusso Kid]

I actually built the whole setup specifically for this. I started with a single GPU, an AMD Ryzen 9, and 64GB of RAM. It worked, but it was pretty limited and struggled to keep up with what I wanted to do.

So I scaled it up. I added another 128GB of RAM and a second GPU, which made a huge difference in performance.

Getting there wasn’t exactly straightforward though. I went through four different motherboards before finding one that could properly support both GPUs. On top of that, I had to upgrade to a 1600 watt power supply and move everything into a much larger case to handle the power and cooling. I now have my eye on the Nvidia 6000 pro's but thats a little bit of an investment, but I would love to see what two of those would do for my ai.

Now it’s finally where I wanted it to be.

I am familiar with the hardware. I built my first system in 2000. Then started building systems for friends and family. For some unknown reason, I have never really lost interest even if I do not build much anymore. The last machine I built was 3 years ago.

There are some that would say you have a bit of an investment in your current set up but jumping it to a couple of 6000 Pros would be huge, in both performance and cost. That would definitely put it into the "investment" field.

 

[That Crusso Kid]

This may sound weird, because I put the shoe on the other foot, but I wish we knew each other better.

 

[mattgsa]

I am familiar with the hardware. I built my first system in 2000. Then started building systems for friends and family. For some unknown reason, I have never really lost interest even if I do not build much anymore. The last machine I built was 3 years ago.

There are some that would say you have a bit of an investment in your current set up but jumping it to a couple of 6000 Pros would be huge, in both performance and cost. That would definitely put it into the "investment" field.

I built my first system in the early 90's, that's also when I started building servers for a small financial company here in San Antonio. My first computer was a AMD 286 16 megahertz, It had a 10mb hard drive and less then a megabyte of RAM, and a monochrome monitory. These day's most IT people don't know what IDE is. Back then that's all that was available. I got out of hardware just before SATA came out, so putting this computer together was interesting to say the least.

 

[That Crusso Kid]

I built my first system in the early 90's, that's also when I started building servers for a small financial company here in San Antonio. My first computer was a AMD 286 16 megahertz, It had a 10mb hard drive and less then a megabyte of RAM, and a monochrome monitory. These day's most IT people don't know what IDE is. Back then that's all that was available. I got out of hardware just before SATA came out, so putting this computer together was interesting to say the least.

I am happy to announce that I do, in fact, know what IDE is. Some of the guys that I rode the train with, around 1994-1998, were programmers and designers.

Since before SATA? That's been a while. You probably had some heart palpatations when seeinng the price tag on the bleeding edge products these days.

 

[mattgsa]

I am happy to announce that I do, in fact, know what IDE is. Some of the guys that I rode the train with, around 1994-1998, were programmers and designers.

Since before SATA? That's been a while. You probably had some heart palpatations when seeinng the price tag on the bleeding edge products these days.

Since I head the IT department, and having owned and MSP, I was familiar with the prices. However, recent events have taken prices to an all-time new high. I'm certainly looking forward to them coming back down. Luckly for me I bought most of my system components like memory and drives just before the prices spiked.

 

[mattgsa]

AI Monitoring Comes to the Fish Room, Meet Jarvise

Most of you know I’ve been messing around with automation and AI for a while now, but I finally started tying it directly into the fish room.

I’ve been building out Jarvise, a fully local AI agent running on my homelab. No cloud services, no subscriptions, no sending data off somewhere else. Everything runs on hardware in my house.

Right now I already have voice and text control working. I can tell Jarvise to turn the lights on or off and it routes the command through Node-RED straight to the Shelly relays. The whole chain is local.

The next step starts tomorrow. I’ve got an Amcrest 4K PoE camera showing up for the fish room. Jarvise is going to start pulling snapshots from the camera on an interval and comparing them against what the system should normally look like.

Things I’m planning on watching for:

• Water level issues in the display or sump
• Cloudy or discolored water
• Fish acting abnormal or hanging at the surface
• Coral looking stressed or retracted across the tank
• Equipment not running correctly — return pump, skimmer, lights, etc.

After watching what happened with POLO Reef, one thing became pretty clear to me, I don’t want to rely on a single notification path or a single controller. If something major happens, I want to know immediately, even if WiFi or one system goes down.

The goal is simple:
If this tank starts dumping water on the floor, I want to know within minutes, not the next morning.

Long term I’ll be adding more sensors into the system too:

• Temp
• pH
• Salinity
• Water level sensors
• Humidity around the sump
• Power monitoring

What I’m really after is correlation between all of it. Not just “a sensor tripped,” but the AI being able to look at camera feeds, environmental data, and equipment status together and recognize when something looks wrong before it becomes a disaster. Or just something as simple as your filter socks are overflowing.

I’m basically trying to build the kind of monitoring system I wish existed for reef tanks already.

More updates once the camera gets integrated and I start training the prompts around what the tank should normally look like.

One of the new custom flush-mount Jarvise control panels is finally coming together.

This panel is going to be the central interface for a lot of the house systems:

• Alarm and security
• Aquarium monitoring and control
• Lighting automation
• Weather and environmental data
• Occupancy awareness (who’s home and where)
• Direct interaction with Jarvise itself

The goal is to make the house feel less like a bunch of disconnected smart devices and more like one integrated system.

Access is controlled through fingerprint authentication, so everyone gets different permissions based on who they are. For example, the kids can control normal house functions, but they won’t be able to mess with the aquarium systems or critical settings.

Everything is running locally on my own hardware, no cloud dependency for core functions. I’m trying to build something that feels more like a real control system than a collection of apps.

Still a lot left to build, but it’s finally starting to feel real.

 

[That Crusso Kid]

AI Monitoring Comes to the Fish Room, Meet Jarvise

Most of you know I’ve been messing around with automation and AI for a while now, but I finally started tying it directly into the fish room.

I’ve been building out Jarvise, a fully local AI agent running on my homelab. No cloud services, no subscriptions, no sending data off somewhere else. Everything runs on hardware in my house.

Right now I already have voice and text control working. I can tell Jarvise to turn the lights on or off and it routes the command through Node-RED straight to the Shelly relays. The whole chain is local.

The next step starts tomorrow. I’ve got an Amcrest 4K PoE camera showing up for the fish room. Jarvise is going to start pulling snapshots from the camera on an interval and comparing them against what the system should normally look like.

Things I’m planning on watching for:

• Water level issues in the display or sump
• Cloudy or discolored water
• Fish acting abnormal or hanging at the surface
• Coral looking stressed or retracted across the tank
• Equipment not running correctly — return pump, skimmer, lights, etc.

After watching what happened with POLO Reef, one thing became pretty clear to me, I don’t want to rely on a single notification path or a single controller. If something major happens, I want to know immediately, even if WiFi or one system goes down.

The goal is simple:
If this tank starts dumping water on the floor, I want to know within minutes, not the next morning.

Long term I’ll be adding more sensors into the system too:

• Temp
• pH
• Salinity
• Water level sensors
• Humidity around the sump
• Power monitoring

What I’m really after is correlation between all of it. Not just “a sensor tripped,” but the AI being able to look at camera feeds, environmental data, and equipment status together and recognize when something looks wrong before it becomes a disaster. Or just something as simple as your filter socks are overflowing.

I’m basically trying to build the kind of monitoring system I wish existed for reef tanks already.

More updates once the camera gets integrated and I start training the prompts around what the tank should normally look like.

One of the new custom flush-mount Jarvise control panels is finally coming together.

This panel is going to be the central interface for a lot of the house systems:

• Alarm and security
• Aquarium monitoring and control
• Lighting automation
• Weather and environmental data
• Occupancy awareness (who’s home and where)
• Direct interaction with Jarvise itself

The goal is to make the house feel less like a bunch of disconnected smart devices and more like one integrated system.

Access is controlled through fingerprint authentication, so everyone gets different permissions based on who they are. For example, the kids can control normal house functions, but they won’t be able to mess with the aquarium systems or critical settings.

Everything is running locally on my own hardware, no cloud dependency for core functions. I’m trying to build something that feels more like a real control system than a collection of apps.

Still a lot left to build, but it’s finally starting to feel real.

Thanks for the update.

You're really starting to pull it together. It's not like I haven't said it before but I'll say it's so very cool again!

I do believe my friend, @tharbin , will find it equally as cool as I do.

 

[That Crusso Kid]

Paging Mr. Harbin. Paging Mr. Tom Harbin... @tharbin

 

[tharbin]

Thanks for the update.

You're really starting to pull it together. It's not like I haven't said it before but I'll say it's so very cool again!

I do believe my friend, @tharbin , will find it equally as cool as I do.

Indeed I do. I've got about 70 pages of reading to do to get the backstory but this post is interesting in itself.

 

[mattgsa]

One thing I’ve realized while building this system is that I’m not really trying to create a “smart aquarium controller” anymore. What I’m actually building is a layered automation and monitoring platform that happens to run a reef tank.

That may sound overkill for a home aquarium, but after seeing enough failures in this hobby — from stuck valves to failed sensors to missed alerts, I started thinking about reef systems more like industrial process systems than hobby equipment.

The biggest shift in my thinking has been this:

AI should not be the first safety layer.

The PLC handles deterministic control and fail-safe logic:

• overflow prevention
• leak response
• pump shutdowns
• interlocks
• emergency states

Jarvise handles the higher-level intelligence:

• telemetry analysis
• trend monitoring
• anomaly detection
• predictive behavior
• long-term learning

The goal is not “AI controls everything.” The goal is teaching the system what normal behavior looks like over time so it can recognize when something no longer makes sense.

For example:

• a valve reports closed but flow never changes
• skimmer production suddenly doubles overnight
• reactor pressure drifts slowly over days
• fish behavior changes before chemistry swings become obvious

Those are the kinds of things humans often notice instinctively, but traditional controllers usually don’t.

That’s the direction I’m trying to push this project toward:
not replacing reef keeping, but building a system that becomes better at understanding the environment over time.

Honestly, even if this never becomes fully autonomous, the process of building it has already changed the way I think about reef automation entirely.