There's a specific kind of frustration that comes from knowing too much about a problem. I had been working as a data engineer for years — building machine learning pipelines, designing data models, thinking about how to make AI actually useful — and every single time I picked up my rod and opened a fishing app, I felt it.
The apps weren't bad. They were just shallow. They showed me wind speed, barometric pressure, water temperature, moon phase. They put pins on maps. They let me log catches. But they didn't understand any of it.
They were showing me raw ingredients and calling it cooking.
The Problem With Every Fishing App I Tried
Picture this: I'm sitting on the bank of a reservoir at 5 AM on a Tuesday in October. The water is 58 degrees. The pressure dropped three hours ago and is now slowly rising. The moon is 70% full and just set. The wind is light out of the northwest. I slept badly. The bass have been hitting topwater in the shallows near submerged timber for the past two weeks — but only at specific times, and only in certain conditions.
I open the fishing app. It tells me the pressure is "good" and that bass prefer 55-65 degree water. Thanks. I knew that.
What I actually needed to know: given all of these specific conditions in combination, at this specific location, with this specific pattern I've been seeing — where should I be right now, and what should I be throwing?
No app could tell me that. Because that question can only be answered by something that can hold all those variables in context simultaneously and cross-reference them against a history of what actually worked. That's not a weather widget problem. That's a machine learning problem.
And that's when I started building Galaxy.
Why Most Apps Get the Architecture Backwards
The standard playbook for building a fishing app goes something like this: design the UI, build the map feature, pull in a weather API, add a catch log, then somewhere down the line — maybe in version 3.0 if the app survives that long — bolt on some AI features.
The AI becomes a layer on top of an architecture that was never designed for it. It's the equivalent of trying to add a foundation to a house after it's already been built. You can do it, but it's clumsy, and the house never sits quite right.
The data you collect, and how you collect it, determines what questions you can ever answer. If you design your data model for display, you'll always be building display features. If you design it for learning, the whole system compounds.
I made a decision early on that felt risky at the time: I wasn't going to build a user interface and figure out AI later. I was going to design the AI infrastructure first — the data model, the AI learning layers, the feedback loops — and then build a user interface that surfaced what the AI understood.
It slowed down early development. There was no v1 to ship in three months. But it meant that every feature we eventually built was standing on a foundation that was actually designed for it.
What "AI-First" Actually Means in Practice
AI-first isn't a marketing term. It's an architectural decision that touches everything.
The data model was designed for machine learning, not display. When you log a catch in Galaxy, you're not just dropping a pin on a map. You're contributing a training example that includes geospatial coordinates, time, weather conditions, water temperature, barometric pressure, moon phase, tide data, your tackle setup, the technique you used, and a dozen other variables — all structured in a way that a model can actually learn from.
Our AI similarity engine lets us ask questions that rules can't answer. This is the part I find most exciting. Galaxy uses high-dimensional embeddings — mathematical representations of fishing conditions that capture the relationship between dozens of variables simultaneously. When you're at a spot and conditions are a certain way, the system can find the most similar historical conditions from across all 40,816 spots in our database and surface what happened then. Not "barometric pressure is rising, so fish shallow" — a rule someone wrote in 2015. Instead: "the last 23 times conditions were this similar to what you're in right now, here's what actually worked."
The algorithm learns from every catch. Galaxy's success scoring model uses a weighted combination of historical success, condition matching, recent community activity, your personal preferences, seasonal patterns, and community validation. Those weights aren't static. Every catch logged — by you, by anyone — feeds back into the model and makes the predictions sharper for everyone who comes after.
Recommendations are pattern-based, not rules-based. There's a profound difference between a system that says "if X then Y" and a system that has learned from millions of data points what actually correlates with success. The former is brittle. It breaks when conditions don't fit the rules. The latter is generative — it can handle combinations of variables that no human ever thought to write a rule for.
The Community Learning Flywheel
Here's something I believe deeply: the best fishing AI in the world, trained on no real fishing data, is worthless. And the most comprehensive fishing database in the world, with no model to learn from it, is just a spreadsheet.
The magic is in the flywheel.
Every catch you log trains the model. A better model makes better predictions. Better predictions help more anglers catch more fish. More fish caught means more catches logged. And the loop continues, compounding with every iteration.
This is why the fishing community isn't just users of Galaxy — they're contributors to something that gets genuinely smarter over time. When an experienced local angler logs a dozen catches on a river system in Montana, they're not just recording their own history. They're teaching the model something about that river, those conditions, that season — knowledge that will eventually help a first-time angler visiting from out of state have a better day on the water.
That's the network effect of building AI-first. The value of the system grows faster than the number of users, because each new user doesn't just add their own data — they improve the predictions for everyone else.
Captain's Mate: Why the AI Layer Had to Come First
When we started building Captain's Mate — Galaxy's LLM-powered fishing guide — a lot of people thought we were building a chatbot. They imagined a customer service bot with some fishing knowledge baked in.
What we actually built is something that could only exist because the AI infrastructure was already there.
Captain's Mate isn't just an LLM with fishing facts in its training data. It's an LLM that has access to Galaxy's entire environmental intelligence layer — the spot histories, the condition embeddings, the species pattern data, the real-time ingestion from your current location. When you ask it a question, it's not retrieving static information. It's reasoning over a live understanding of current conditions, your personal catch history, and everything the model has learned from the broader Galaxy community.
That only works if you built the AI layer first. If the intelligence lives in the model, the guide can tap into it. If the intelligence lives in a rule set someone wrote, the guide can only recite rules.
The difference matters enormously when you're standing in front of a new body of water at dawn and you need real guidance, not a pamphlet.
What an AI Can Do That No Human Guide Can
I want to be careful here, because I have enormous respect for experienced fishing guides. The best ones carry decades of embodied knowledge about reading water, understanding fish behavior, and adapting to conditions in real time. That knowledge is irreplaceable.
But there are specific things an AI can do that no human can:
- Process the catch histories of 40,816 distinct fishing spots simultaneously when generating a recommendation
- Cross-reference millions of condition-catch data points in under a second
- Identify correlations between variables that are too subtle or too multi-dimensional for human pattern recognition to detect
- Maintain perfect recall of every condition that preceded every successful catch, with zero nostalgia bias or selective memory
- Apply what worked in one corner of the world to a question being asked from another corner, the moment the patterns are similar enough to be relevant
The best fishing guide you've ever had knows their home water deeply. Galaxy knows every water in our database with the same depth — and it updates that knowledge in real time.
The Honest Limitations
I'd be doing you a disservice if I didn't say this plainly: AI doesn't make you a better angler by itself.
Galaxy can tell you the right spot. It can tell you the right time. It can surface the patterns that suggest topwater over a particular piece of structure in the next 45 minutes. But if you can't make an accurate cast, if you set the hook too early, if your retrieve speed is off — the fish don't care how sophisticated the algorithm is.
What Galaxy gives you is better information, so you can make better decisions. The fishing still happens between you and the fish. The reading of water, the feel for what a lure is doing, the intuition built over years of time on the water — those are human things. They should stay human things.
Think of it the way a pilot thinks about instruments. The instruments don't fly the plane. But a pilot who ignores them isn't more skilled — they're just flying blind.
The Next Three Years
I'll tell you where we're headed, because I think it's worth saying out loud.
Predictive species migration mapping. We're working on models that can predict where fish populations will be — not just where they are now — based on temperature trends, bait movement, and seasonal patterns. Think of it as a weather forecast for fish.
Personalized bite windows. The community model tells us when fish are active in aggregate. But your catch history tells us something more specific: when you tend to catch fish, with your style, your tackle, your technique. We're building personalization that goes all the way down to your individual patterns.
AI-powered tackle recommendations. Based on current conditions, current fish behavior, and what's actually working in the community right now — not what the tackle company paid to promote. Real recommendations, grounded in data.
The thing that makes all of this possible is the foundation. Because we built AI-first, every one of these features has somewhere to live. The data is structured for it. The models are learning toward it. The infrastructure is ready.
Why I'm Sharing This
I'm writing this because I want the Galaxy community to understand what they're part of. When you log a catch, you're not just keeping a fishing journal. You're contributing to a model that helps anglers across the country catch more fish. You're part of a flywheel that, frankly, I think is going to change what fishing apps are capable of.
And I'm writing this because I want other anglers who are also engineers, or engineers who also fish, to know that there's a different way to build tools for the sports we love. The category of "fishing apps" doesn't have to be weather widgets with map pins. It can be genuinely intelligent.
If you're not on Galaxy yet, download the app and start logging your catches. Every entry makes the AI smarter — for you and for everyone else. If you have questions about how any of this works, Captain's Mate is always available in the app, and I read every message that comes through our community channels.
The water's out there. Let's find the fish.
— Levi, Founder of Galaxy Fishing