Hardware & Software Overhaul: Using Gemini as my "Lead Engineer."

[Xyloxo]

Hey everyone,
I’m currently in the middle of a major hardware and software overhaul (controller, coding, and cabinet redesign), and I wanted to start a discussion about a tool I’m leaning on heavily during this update: Google's Gemini.
First off, Reef2Reef is hands down the best source of community and inspiration out there. I come here because the human element is irreplaceable. But we all know the struggle: you ask a technical question about a build, and you get 15 different answers, 3 arguments, and a lot of brand bias.
That’s where I’m finding Gemini useful right now. It doesn't have an ego, it’s not loyal to a specific brand, and it gives me objective, unbiased data to help me make decisions as I upgrade the system.
Here is a look at what I am working on right now:
1. "Visual" Engineering (The Current Phase)
This has been the coolest part of the upgrade so far. Instead of just typing out lists, I’m uploading photos of my equipment piles to Gemini.
Inventory & Design: I use it to scan my photos to identify the specific modules, pumps, and lights I have on hand.
Layout: I am currently using those visuals to design the physical layout of the new controller board. Since it "sees" the size of the power bricks and modules, it helps me arrange them to fit perfectly.
Custom Code: I’m getting it to write the Neptune Apex code based specifically on the hardware it identified in the pictures, rather than just using generic templates.
2. Dialing in the Bacteria Protocol
I use a combination of MicroBacter 7, MicroBacter Clean, and TLC Marine SAT.
The Logic: Since I dose these by hand, I needed a way to make sure I didn't mess up the staggering. I just finished building a weekly calendar with Gemini to ensure I’m not overlapping products inappropriately.
The Automation: I uploaded my current Apex code to see if it could handle the routine. Gemini analyzed it, corrected my syntax errors, and suggested specific additions (and removals) to make the logic cleaner. The result is a streamlined Virtual Outlet: when I trigger "Dosing Mode," it kills the Skimmer, UV, and Return Pump, waits 4 hours, and then brings them back online in a staggered sequence so the skimmer doesn't go crazy. It also forces my Auto Water Change to skip if it overlaps.
3. Engineering the Electrical Topology
I’m still finalizing the controller cabinet, and safety is the main priority.
Power Architecture: I am mapping out the safest path: Wall > GFCI > Pure Sine Wave UPS > Heavy Duty Trip-Lite Strip > Dual Neptune EB8s.
Load Calcs: I am actively calculating the amperage draw of the heaters and pumps to ensure the UPS is sized correctly for the load.
Interference: My current plan involves using slotted finger ducts to strictly separate the high-voltage 120v lines from the sensitive probe data lines to avoid the interference issues I’ve had in the past.
4. The "Myriad" of Troubleshooting
Beyond the build, it’s become my go-to for the daily hiccups that pop up while getting a tank running. It’s helping me with:
Flow Dynamics: Describing dead spots and working on nozzle orientation.
Noise Hunting: Tracking down vibration sources in the plumbing.
Parameter Trends: I’m feeding it test results to spot consumption trends before they become problems.
The Question:
This project feels different because I have a "reefing engineer" on call 24/7 to double-check my math and my logic.
Is anyone else building their tank this way?
Are you using AI to help design your plumbing or electrical layouts?
Have you trusted it to help write your controller code yet?
Or do you prefer to stick to the forum consensus and your own experience?

 

[lapin]

I used 1” thick acrylic and 2 part glue. Fastened wires with clips and stainless screws.
Don’t need rocket science to run a reef tank.
Seems a bit odd. Gemini can’t clean your glass or syphon the sand bed.

 

[BeanAnimal]

A few notes:

Gemeni, GPT, etc, don’t keep context well. They can write code, but also make a mess of code. Fix one thing and break another, or decide to “optimize” things that shouldn’t be touched. They take liberties even when you ask them not to, and can’t remember you told them not to. So you must not trust anything and must check everything, every prompt every time.

They are not engineers, nor do they apply logic. They tell you what you want to hear based on how you ask. Don’t mistake that for intelligence or understanding. They don’t have an ego, but they don’t know how to say no, to the point of fabricating something that sounds good vs it actually making sense.

They will happily spout bad math, misapplied formulas or logic, or simply make up science, and may do it at any time, even amidst other correct information. They are know it all’s that actually know nothing other than language and what words commonly fit and appear together.

 

[rtparty]

Be very careful with Apex code and AI. I had someone message me asking why their Apex wasn’t working properly and when I saw the code it had contradictions all over the place. Asked them who coded it and they told me ChatGPT.

I’m not even good with Apex and code but I still saw the issues

 

[Xyloxo]

Thanks for the feedback, everyone. These are all really valid points, and I think it highlights exactly how this tool should be used versus how it can go wrong.

To address a few of the specific concerns:

"Gemini can't clean your glass."
100% agree. Nothing replaces wet hands in the tank. The way I look at it is: the AI handles the "mental load"—the math, the calendar planning, the code syntax—so I have more time and energy to focus on the actual husbandry, like water changes and glass cleaning. It doesn't replace the reefer; it just organizes the reefer's to-do list.

"AI makes a mess of code / hallucinates math / loses context."
You are absolutely right, and this is the most important rule of using AI: Trust but Verify.
I treat Gemini like a "Junior Intern." It does the heavy lifting on writing the initial draft of the code or doing the rough math, but I am the Senior Engineer who has to review and approve it.
I don't blindly paste the code. I read it to ensure the logic flows (e.g., checking that Defer statements are in the right order).
I don't blindly trust the dosage math. I ask it to "show its work" so I can spot if it used the wrong formula.

"It's not an engineer."
True. It’s a language model. But it is a language model that has read the entire Neptune Systems manual and forums. When I asked it to design the cabinet layout, I didn't just ask "make it cool." I gave it the specific dimensions of my gear and asked for a layout that separates high voltage from data lines. It gave me a great starting point that I could then tweak.

The Bottom Line:

It’s a powerful tool, but dangerous if you use it on "Autopilot." If you use it to assist your own knowledge, it's incredible. If you use it to replace having to learn how your tank works, you're going to crash your tank.

 

[BeanAnimal]

It may have read the Neptune manual, but it is still and still LLM and will happily conflate that information with anything else it has read.

Line voltage and data lines don’t necessarily need to be separated, it depends on their function and shielding. The LLM doesn’t know or understand this, it is repeating words.

Asking it to show its work means you have to check the math, it will happily add 2+2 to get five and verify that 5 is the answer.

But, more importantly; Please don’t use AI to generate responses here. You asked it to help you respond, as clearly evident in the language and structure of your last response. To be sure, we prefer talking to people and interacting with people, not AI through human proxies or AI derived language, arguments or barely tangentially relevant AI talking points.

 

[fbg]

Software engineer here. I agree with BeanAnimal in general, but there have been substantial improvements in LLM performance (hallucinations, context, sycophancy, etc) in the last year.

My colleagues and I routinely use a LLM (Claude Code) for software development in the medical field. In good hands, it is a tremendous help. I personally do not know any professional in software engineering who is not taking advantage of LLMs to some extent. These tools are not perfect (far from there), and indeed we still do a full code review, but the potential is clearly huge.

I totally agree that in bad hands (vibe coding with no quality control), LLMs could be a recipe for disaster.

There are special techniques (e.g. RAG, Retrieval-Augmented Generation) to ensure that the output respects closely a certain knowledge base, for example the Neptune manuals. In my field, there are several projects using RAG or similar to do clinical decision support/treatment recommendations - quite delicate applications. Of course not easy, and the regulation is strict (at least here in Europe), but there are some products already hitting the market.

 

[Subsea]

Hey everyone,
I’m currently in the middle of a major hardware and software overhaul (controller, coding, and cabinet redesign), and I wanted to start a discussion about a tool I’m leaning on heavily during this update: Google's Gemini.
First off, Reef2Reef is hands down the best source of community and inspiration out there. I come here because the human element is irreplaceable. But we all know the struggle: you ask a technical question about a build, and you get 15 different answers, 3 arguments, and a lot of brand bias.
That’s where I’m finding Gemini useful right now. It doesn't have an ego, it’s not loyal to a specific brand, and it gives me objective, unbiased data to help me make decisions as I upgrade the system.
Here is a look at what I am working on right now:
1. "Visual" Engineering (The Current Phase)
This has been the coolest part of the upgrade so far. Instead of just typing out lists, I’m uploading photos of my equipment piles to Gemini.
Inventory & Design: I use it to scan my photos to identify the specific modules, pumps, and lights I have on hand.
Layout: I am currently using those visuals to design the physical layout of the new controller board. Since it "sees" the size of the power bricks and modules, it helps me arrange them to fit perfectly.
Custom Code: I’m getting it to write the Neptune Apex code based specifically on the hardware it identified in the pictures, rather than just using generic templates.
2. Dialing in the Bacteria Protocol
I use a combination of MicroBacter 7, MicroBacter Clean, and TLC Marine SAT.
The Logic: Since I dose these by hand, I needed a way to make sure I didn't mess up the staggering. I just finished building a weekly calendar with Gemini to ensure I’m not overlapping products inappropriately.
The Automation: I uploaded my current Apex code to see if it could handle the routine. Gemini analyzed it, corrected my syntax errors, and suggested specific additions (and removals) to make the logic cleaner. The result is a streamlined Virtual Outlet: when I trigger "Dosing Mode," it kills the Skimmer, UV, and Return Pump, waits 4 hours, and then brings them back online in a staggered sequence so the skimmer doesn't go crazy. It also forces my Auto Water Change to skip if it overlaps.
3. Engineering the Electrical Topology
I’m still finalizing the controller cabinet, and safety is the main priority.
Power Architecture: I am mapping out the safest path: Wall > GFCI > Pure Sine Wave UPS > Heavy Duty Trip-Lite Strip > Dual Neptune EB8s.
Load Calcs: I am actively calculating the amperage draw of the heaters and pumps to ensure the UPS is sized correctly for the load.
Interference: My current plan involves using slotted finger ducts to strictly separate the high-voltage 120v lines from the sensitive probe data lines to avoid the interference issues I’ve had in the past.
4. The "Myriad" of Troubleshooting
Beyond the build, it’s become my go-to for the daily hiccups that pop up while getting a tank running. It’s helping me with:
Flow Dynamics: Describing dead spots and working on nozzle orientation.
Noise Hunting: Tracking down vibration sources in the plumbing.
Parameter Trends: I’m feeding it test results to spot consumption trends before they become problems.
The Question:
This project feels different because I have a "reefing engineer" on call 24/7 to double-check my math and my logic.
Is anyone else building their tank this way?
Are you using AI to help design your plumbing or electrical layouts?
Have you trusted it to help write your controller code yet?
Or do you prefer to stick to the forum consensus and your own experience?

As a controls systems engineer in deep water drilling, I am old school. Not much interested in ChatGPT running my reef aquarium control system architecture.

 

[BeanAnimal]

Software engineer here. I agree with BeanAnimal in general, but there have been substantial improvements in LLM performance (hallucinations, context, sycophancy, etc) in the last year.

I use both GPT and Claude in the software engineering space as well. There is a stark difference between code assistance for a software engineer and the AI being a personal reefing assistant or authoritative source of information. Code documentation, also hit or miss without rigid framework.

In the cases of the latest Gemini and GPT the sycophancy is worse… far worse if you ask me. Feels like they put a lot of effort into emojis and platitudes though.

In any case, context here is the use case presented for reefing by the average user. Not templated and scoped use in a structured environment designed to leverage the tool and account for its limitations and flaws.

 

[fbg]

In any case, context here is the use case presented for reefing by the average user. Not templated and scoped use in a structured environment designed to leverage the tool and account for its limitations and flaws.

Yes, indeed.

Off-topic, for curiosity only - our scenario here reminds me of patients without specific medical knowledge asking LLMs information about their pathologies, treatments etc. This use case has been deeply investigated in scientific literature, usually by asking a pool of human experts to evaluate in a structured way the replies to a set of questions provided by the LLMs (typically comparing GPT with Gemini etc). Results tend to be generally positive but with several exceptions, with an overall tendency towards improvement with the newer versions. It might be more or less the same for reefing - but with a much smaller body of literature in the training data, thus who knows.

 

[Randy Holmes-Farley]

Yes, indeed.

Off-topic, for curiosity only - our scenario here reminds me of patients without specific medical knowledge asking LLMs information about their pathologies, treatments etc. This use case has been deeply investigated in scientific literature, usually by asking a pool of human experts to evaluate in a structured way the replies to a set of questions provided by the LLMs (typically comparing GPT with Gemini etc). Results tend to be generally positive but with several exceptions, with an overall tendency towards improvement with the newer versions. It might be more or less the same for reefing - but with a much smaller body of literature in the training data, thus who knows.

One of the clear problems with a reefing AI question as opposed to a properly trained medical AI is that the general AI's train on all internet stuff, which for reefing is often inaccurate and is simply repeated misinformation that appears over and over. It is also misled by product claims, which AI seem to often just accept.

 

[fbg]

One of the clear problems with a reefing AI question as opposed to a properly trained medical AI is that the general AI's train on all internet stuff, which for reefing is often inaccurate and is simply repeated misinformation that appears over and over. It is also misled by product claims, which AI seem to often just accept.

I guess we'll soon see a commercial AI-powered reefing assistant. I'm betting on Red Sea.

 

[BeanAnimal]

What everyone needs, an AI driven marketing tool right in their home that will guarantee success if they just follow directions and buy the products being peddled.

Which, speaking of GPT, has touted new tool; Shopping Research (or whatever they labeled it) that is certainly already being gamed and the new wild west of black hat marketing.