2026-02-23, 04:26 PM
THIS IS NOT AN OPENPLOTTER REPLACEMENT. Open Plotter is Brilliant!
I used AI Assistant to create something can be accompanied to Open Plotter or any third party Plotter or Fish Finder.
A boater, a dream, and four AI tools walk into a code editor...
I'll be honest with you. When I started building d3kOS, I didn't fully know what I was building. I knew the problem: I had a boat with analogue engine gauges — the kind that have been on vessels since your grandfather's day — and I wanted to drag them kicking and screaming into the age of artificial intelligence. RPM, oil pressure, coolant temperature, fuel levels — all of it sitting there in perfectly good analogue gauges, completely invisible to any modern diagnostic system. it running on Jellyfish box.
The vision was clear. The path to get there? Less so.
What followed was one of the most educational, frustrating, exciting, and occasionally maddening experiences of my development life. And I didn't do it alone. I did it with four AI tools, a Raspberry Pi 4B, and more coffee than I care to admit.
The Problem Worth Solving
For boaters with older vessels, the engine monitoring situation is a tale of two worlds. Big money boats run PTZ cameras, full NMEA2000 networks, and chartplotters that would look at home on the bridge of a container ship. The rest of us have gauges. Good gauges. Reliable gauges. But dumb gauges.
The CX5106 NMEA2000 gateway changes that — it connects to your existing analogue gauge senders and converts those signals to digital NMEA2000 data that modern electronics can actually read. That was the hardware foundation. What I needed was the software brain to sit on top of it.
I wanted real-time monitoring. Analytics. Anomaly detection. And the thing that really fired my imagination: an AI assistant that could read your actual engine manual as a PDF, cross-reference it with live sensor data, and tell you why your oil pressure was doing something weird at 2,000 RPM.
That's d3kOS. Marine Intelligence Operating System. And building it meant figuring out which AI tools could actually help me build it — and which ones were going to waste my time.
The AI Lineup: Four Tools, One Project
ChatGPT — The Brainstormer
ChatGPT was where a lot of ideas started. It's excellent at the early-stage stuff: "Help me think through the architecture for a marine monitoring system." "What are the pros and cons of using Signal K versus building my own data layer?" "Walk me through how NMEA2000 data flows."
For high-level conceptual work, for getting unstuck when you're staring at a blank page, ChatGPT delivered. It's like having a very well-read colleague who's happy to think out loud with you at any hour.
Where it started to show its limits was in the details. Ask it to help you build something specific — a particular API endpoint, a WebSocket handler for live engine data — and the code would often be plausible-looking but subtly wrong. You'd implement it, it wouldn't work, you'd ask why, and the conversation would spiral. It knew a lot, but it didn't always know what it didn't know.
Perplexity — The Researcher
Perplexity became my go-to for research that needed to be grounded in reality. "What's the current state of YOLOv8 for embedded systems?" "What are boaters actually using for onboard NVR solutions?" "What does the Signal K spec say about engine data PGNs?"
It's a researcher's tool — fast, sourced, current. When I needed to understand something that existed out in the world rather than something I needed to create, Perplexity was the first window I opened. For d3kOS specifically, it helped me understand the NMEA2000 ecosystem, evaluate camera options for the Marine Vision module, and get up to speed quickly on PocketSphinx and Vosk for the voice command system.
But Perplexity doesn't write your code. It informs the decisions that lead to the code. It's a brilliant research assistant, not a development partner.
GitHub Copilot — The Autocomplete Artist
Copilot is a different beast entirely — it lives inside your editor, watching what you type and whispering suggestions. For certain kinds of work it's genuinely magical. Repetitive patterns, boilerplate, standard implementations — Copilot handles these with the confidence of someone who has read every Stack Overflow answer ever written.
Building out the twelve microservices that power d3kOS involved a lot of that kind of work. API endpoints follow patterns. Data structures repeat. Copilot would often complete an entire function before I'd finished typing the first line, and it would be correct.
The challenge came with the novel stuff. d3kOS isn't just CRUD operations and standard patterns — it's a self-healing system with voice wake words, an onboard AI that reads marine manuals, a touchscreen interface designed for wet hands in rough conditions. When the problem was genuinely new, Copilot's suggestions became less reliable. It was autocompleting from patterns that didn't quite fit, and I was spending more time rejecting suggestions than accepting them.
Claude Code — The One That Changed Everything
I came to Claude Code later in the process, and I'll admit I came to it with some skepticism. I'd already been burned a few times by AI-generated code that looked right and wasn't. I was cautious.
The first thing I noticed was that Claude Code actually read the context. Not just the file I was working in — the whole project. It understood that the voice system and the touchscreen interface shared a conflict. It flagged it. "If you stop the voice service this way, it will break the touchscreen — here's why, and here's how to handle it." That kind of systems-level awareness was something I hadn't experienced with the other tools.
For d3kOS specifically, Claude Code became indispensable for:
The complex integrations. Getting the CX5106 talking to the PiCAN-M HAT talking to Signal K talking to the dashboard — that's not a simple chain. Claude Code held the whole architecture in mind and helped debug across layers in a way that felt like pairing with a senior engineer rather than querying a search engine.
The AI assistant module itself. The irony of using AI to build an AI assistant isn't lost on me. The onboard diagnostic system — the part that reads your engine manual and cross-references it with live sensor data — required careful design. Claude Code helped architect the dual-mode system (online via OpenRouter, offline with cached instant-answer patterns) and got the response caching working to the point where common queries now return in 0.17 seconds instead of eighteen.
The self-healing system* Twelve microservices on a Raspberry Pi, all monitoring each other, automatically restarting on failure. The watchdog logic, the health checks, the thermal protection — Claude Code helped build a system that genuinely recovers from its own failures. On a boat, far from shore, that matters enormously.
The touchscreen interface. Designing UI for wet hands and rough conditions is a specific discipline. d3kOS runs on a 10.1" touchscreen — large enough to read at a glance from the helm, even in bright sunlight or at night. That screen real estate informed every design decision: 60-pixel minimum button targets so you can hit them with a wet thumb in a rolling sea, 48-pixel gauge values readable from across the cockpit, a dark theme optimized for night vision, and an on-screen keyboard that activates automatically so you never need to hunt for a physical input. Claude Code understood the constraints and helped enforce them consistently across every page.
What I Learned About AI-Assisted Development
After all of it — the late nights, the crashed builds, the moments of breakthrough — here's what I'd tell anyone embarking on a serious AI-assisted development project:
Use each tool for what it's actually good at.** Perplexity for research. ChatGPT for brainstorming and architecture. Copilot for pattern-heavy implementation work. Claude Code for complex, interconnected problems where context and systems thinking matter.
The right AI doesn't just answer your question — it improves your question.** The best interactions I had with Claude Code weren't ones where I asked something and got an answer. They were ones where it pushed back, asked what I was actually trying to achieve, and helped me realize I was solving the wrong problem.
AI accelerates development but doesn't replace judgment. d3kOS is a beta product — deliberately, carefully labeled as not suitable for production marine use. The AI tools helped me build faster than I could have alone, but every safety decision, every architectural tradeoff, every "this feature isn't ready yet" call — those were mine. The AI is a collaborator, not a captain.
Backups. Always backups. (Ask me about my last 48 hours sometime.)
Where d3kOS Stands Today
Version 2.0-T3 is out as of February 2026. It's a testing build — real, functional, running on Raspberry Pi 4B with a 10.1" touchscreen, reading analogue gauges through the CX5106, displaying live engine data, answering voice queries, streaming camera feeds, detecting objects with YOLOv8.
Fish identification is coming in Phase 2.2. The mobile app is on the roadmap. A billing system is in development. There's a WiFi hotspot limitation I'm still fighting with BCM4345/6 firmware.
It's a boat project. There's always something.
But the core of it — an open-source marine intelligence system that any boater with an analogue dashboard can install and actually use — that part works. And it exists, in large part, because four AI tools showed up and helped me build it. One of them more than the others.
d3kOS is open source and available on GitHub. Beta testers welcome.
- GitHub: [github.com/SkipperDon/d3kOS](https://github.com/SkipperDon/d3kOS)
- Download: v2.0-T3 (9.2 GB image)
- Support: support@atmyboat.com
- Website: [atmyboat.com](https://atmyboat.com)
Skipper Don is the developer of d3kOS and writes about marine technology, AI-assisted development, and the eternal optimism required to maintain an older boat.
I used AI Assistant to create something can be accompanied to Open Plotter or any third party Plotter or Fish Finder.
A boater, a dream, and four AI tools walk into a code editor...
I'll be honest with you. When I started building d3kOS, I didn't fully know what I was building. I knew the problem: I had a boat with analogue engine gauges — the kind that have been on vessels since your grandfather's day — and I wanted to drag them kicking and screaming into the age of artificial intelligence. RPM, oil pressure, coolant temperature, fuel levels — all of it sitting there in perfectly good analogue gauges, completely invisible to any modern diagnostic system. it running on Jellyfish box.
The vision was clear. The path to get there? Less so.
What followed was one of the most educational, frustrating, exciting, and occasionally maddening experiences of my development life. And I didn't do it alone. I did it with four AI tools, a Raspberry Pi 4B, and more coffee than I care to admit.
The Problem Worth Solving
For boaters with older vessels, the engine monitoring situation is a tale of two worlds. Big money boats run PTZ cameras, full NMEA2000 networks, and chartplotters that would look at home on the bridge of a container ship. The rest of us have gauges. Good gauges. Reliable gauges. But dumb gauges.
The CX5106 NMEA2000 gateway changes that — it connects to your existing analogue gauge senders and converts those signals to digital NMEA2000 data that modern electronics can actually read. That was the hardware foundation. What I needed was the software brain to sit on top of it.
I wanted real-time monitoring. Analytics. Anomaly detection. And the thing that really fired my imagination: an AI assistant that could read your actual engine manual as a PDF, cross-reference it with live sensor data, and tell you why your oil pressure was doing something weird at 2,000 RPM.
That's d3kOS. Marine Intelligence Operating System. And building it meant figuring out which AI tools could actually help me build it — and which ones were going to waste my time.
The AI Lineup: Four Tools, One Project
ChatGPT — The Brainstormer
ChatGPT was where a lot of ideas started. It's excellent at the early-stage stuff: "Help me think through the architecture for a marine monitoring system." "What are the pros and cons of using Signal K versus building my own data layer?" "Walk me through how NMEA2000 data flows."
For high-level conceptual work, for getting unstuck when you're staring at a blank page, ChatGPT delivered. It's like having a very well-read colleague who's happy to think out loud with you at any hour.
Where it started to show its limits was in the details. Ask it to help you build something specific — a particular API endpoint, a WebSocket handler for live engine data — and the code would often be plausible-looking but subtly wrong. You'd implement it, it wouldn't work, you'd ask why, and the conversation would spiral. It knew a lot, but it didn't always know what it didn't know.
Perplexity — The Researcher
Perplexity became my go-to for research that needed to be grounded in reality. "What's the current state of YOLOv8 for embedded systems?" "What are boaters actually using for onboard NVR solutions?" "What does the Signal K spec say about engine data PGNs?"
It's a researcher's tool — fast, sourced, current. When I needed to understand something that existed out in the world rather than something I needed to create, Perplexity was the first window I opened. For d3kOS specifically, it helped me understand the NMEA2000 ecosystem, evaluate camera options for the Marine Vision module, and get up to speed quickly on PocketSphinx and Vosk for the voice command system.
But Perplexity doesn't write your code. It informs the decisions that lead to the code. It's a brilliant research assistant, not a development partner.
GitHub Copilot — The Autocomplete Artist
Copilot is a different beast entirely — it lives inside your editor, watching what you type and whispering suggestions. For certain kinds of work it's genuinely magical. Repetitive patterns, boilerplate, standard implementations — Copilot handles these with the confidence of someone who has read every Stack Overflow answer ever written.
Building out the twelve microservices that power d3kOS involved a lot of that kind of work. API endpoints follow patterns. Data structures repeat. Copilot would often complete an entire function before I'd finished typing the first line, and it would be correct.
The challenge came with the novel stuff. d3kOS isn't just CRUD operations and standard patterns — it's a self-healing system with voice wake words, an onboard AI that reads marine manuals, a touchscreen interface designed for wet hands in rough conditions. When the problem was genuinely new, Copilot's suggestions became less reliable. It was autocompleting from patterns that didn't quite fit, and I was spending more time rejecting suggestions than accepting them.
Claude Code — The One That Changed Everything
I came to Claude Code later in the process, and I'll admit I came to it with some skepticism. I'd already been burned a few times by AI-generated code that looked right and wasn't. I was cautious.
The first thing I noticed was that Claude Code actually read the context. Not just the file I was working in — the whole project. It understood that the voice system and the touchscreen interface shared a conflict. It flagged it. "If you stop the voice service this way, it will break the touchscreen — here's why, and here's how to handle it." That kind of systems-level awareness was something I hadn't experienced with the other tools.
For d3kOS specifically, Claude Code became indispensable for:
The complex integrations. Getting the CX5106 talking to the PiCAN-M HAT talking to Signal K talking to the dashboard — that's not a simple chain. Claude Code held the whole architecture in mind and helped debug across layers in a way that felt like pairing with a senior engineer rather than querying a search engine.
The AI assistant module itself. The irony of using AI to build an AI assistant isn't lost on me. The onboard diagnostic system — the part that reads your engine manual and cross-references it with live sensor data — required careful design. Claude Code helped architect the dual-mode system (online via OpenRouter, offline with cached instant-answer patterns) and got the response caching working to the point where common queries now return in 0.17 seconds instead of eighteen.
The self-healing system* Twelve microservices on a Raspberry Pi, all monitoring each other, automatically restarting on failure. The watchdog logic, the health checks, the thermal protection — Claude Code helped build a system that genuinely recovers from its own failures. On a boat, far from shore, that matters enormously.
The touchscreen interface. Designing UI for wet hands and rough conditions is a specific discipline. d3kOS runs on a 10.1" touchscreen — large enough to read at a glance from the helm, even in bright sunlight or at night. That screen real estate informed every design decision: 60-pixel minimum button targets so you can hit them with a wet thumb in a rolling sea, 48-pixel gauge values readable from across the cockpit, a dark theme optimized for night vision, and an on-screen keyboard that activates automatically so you never need to hunt for a physical input. Claude Code understood the constraints and helped enforce them consistently across every page.
What I Learned About AI-Assisted Development
After all of it — the late nights, the crashed builds, the moments of breakthrough — here's what I'd tell anyone embarking on a serious AI-assisted development project:
Use each tool for what it's actually good at.** Perplexity for research. ChatGPT for brainstorming and architecture. Copilot for pattern-heavy implementation work. Claude Code for complex, interconnected problems where context and systems thinking matter.
The right AI doesn't just answer your question — it improves your question.** The best interactions I had with Claude Code weren't ones where I asked something and got an answer. They were ones where it pushed back, asked what I was actually trying to achieve, and helped me realize I was solving the wrong problem.
AI accelerates development but doesn't replace judgment. d3kOS is a beta product — deliberately, carefully labeled as not suitable for production marine use. The AI tools helped me build faster than I could have alone, but every safety decision, every architectural tradeoff, every "this feature isn't ready yet" call — those were mine. The AI is a collaborator, not a captain.
Backups. Always backups. (Ask me about my last 48 hours sometime.)
Where d3kOS Stands Today
Version 2.0-T3 is out as of February 2026. It's a testing build — real, functional, running on Raspberry Pi 4B with a 10.1" touchscreen, reading analogue gauges through the CX5106, displaying live engine data, answering voice queries, streaming camera feeds, detecting objects with YOLOv8.
Fish identification is coming in Phase 2.2. The mobile app is on the roadmap. A billing system is in development. There's a WiFi hotspot limitation I'm still fighting with BCM4345/6 firmware.
It's a boat project. There's always something.
But the core of it — an open-source marine intelligence system that any boater with an analogue dashboard can install and actually use — that part works. And it exists, in large part, because four AI tools showed up and helped me build it. One of them more than the others.
d3kOS is open source and available on GitHub. Beta testers welcome.
- GitHub: [github.com/SkipperDon/d3kOS](https://github.com/SkipperDon/d3kOS)
- Download: v2.0-T3 (9.2 GB image)
- Support: support@atmyboat.com
- Website: [atmyboat.com](https://atmyboat.com)
Skipper Don is the developer of d3kOS and writes about marine technology, AI-assisted development, and the eternal optimism required to maintain an older boat.
Skipper Don
AtMyBoat.com
skipperdon@atmyboat.com
I watched enough Gilligan’s Island as a kid to explain this
AtMyBoat.com
skipperdon@atmyboat.com
I watched enough Gilligan’s Island as a kid to explain this