Route Engine
vs. Route Engine 2.0.

We are building a second routing engine. That raises a fair question: what is wrong with the first one? The short answer is nothing serious. The longer answer is what this post is about.

The problem, in one sentence

Every shipment on the platform needs a path across the ocean. One that does not cut through an island, run a boat aground in shallow water, or take a wildly longer path than necessary. Working out that path automatically, for any two ports in Indonesia and beyond, is what the Route Engine does. Both versions of it.

The simple picture first

Imagine the ocean as a big blue puzzle full of islands. Rocks poking out of a lake. A toy boat needs to sail from one dock to another without hitting any of them.

To find that path, the computer does not just draw a straight line. It places a set of stepping stones into the water first, then hops from stone to stone to find a safe way across. Like a game of hopscotch that always avoids the rocks.

That is really all routing is: placing stepping stones, then finding the best path through them. The difference between our two engines comes down to when and how those stepping stones get placed.

The Route Engine today

Every time someone clicks auto-generate on a route, the engine works on the spot. It grabs the coastline map for that area, quickly scatters a fresh handful of stepping stones just for that one trip, and hops through them to find a path. All in a few seconds.

It is fast, and over this year it has gotten genuinely good. A few real numbers sit behind that claim.

• It checks real water depth along the path, so a route does not send a ship somewhere it would run aground.
• It has learned eight specific real shortcuts: actual rivers, bays, and harbor channels that a general map is too blurry to see on its own, hand-verified with real coastline data. The Siak River, Port Klang's north channel, and Kendari Bay are among them.
• A systematic check across every close pair of ports in the system found it was taking absurd detours in dozens of cases. One route between two ports just 4.5 nautical miles apart was drawing a 227-mile loop. Every one of those bugs has since been found and fixed, with real before-and-after numbers to prove it.

Route Engine 2.0

Instead of scattering stepping stones fresh every time a route is requested, Route Engine 2.0 sits down ahead of time and carefully places a very large number of stones across the entire map. Thousands of them, packed extra close together near tricky spots like narrow straits, and spread wider out in open ocean where there is nothing to hit. It saves that whole map of stones once, offline, in about thirty minutes.

When someone asks for a route, nothing new gets built. The engine just finds the best path through the stones that are already sitting there. Because there are so many more stones, placed more thoughtfully, the path tends to look smoother. And it can be found faster, especially as the platform handles more and more requests at once.

For the full Indonesian archipelago, that comes out to just over 24,000 cells, more than 222,000 connections between them, and 266 cells classified as genuine strait passages, where a real check confirmed coastline on two opposing sides rather than just somewhere near land.

Right now, Route Engine 2.0 is running on a private diagnostic page only our team can see. It is not yet the thing that plans real shipments, and we would rather say that plainly than imply otherwise.

The machine learning part is still learning

We call this smart routing and we mean it, but it is worth being honest about where the learning actually stands today.

The system uses GPS trails from real voyages your fleet has already completed to make future route suggestions more accurate than the ones before. That is machine learning in the genuine sense of the phrase. The more your fleet sails, the more the system has to work with, and the better its suggestions get for your specific lanes, not anyone else's.

Right now, for a fleet early in its time on the platform, the suggestions lean more heavily on depth, coastline, and corridor data than on accumulated voyage history. That is not a bug. It is just an honest reflection of where the learning is. A fleet that has completed ten voyages gets a better suggestion than a fleet that has completed one, and a fleet that has completed a hundred gets a better suggestion still.

We are not pretending the system arrives fully formed. It does not. It arrives capable, and it gets more accurate over time as it learns from real sailings. That is the point of building it this way.

So is the current engine not good enough?

It is genuinely solid. The Route Engine handles our real, day-to-day maritime route generation. It checks depth safety. It knows real shortcuts through rivers and bays that took real investigation to find. Every disproportionate-detour bug we could locate across the whole system has been tracked down and fixed with real evidence, not guesswork.

But there are a few things it will always struggle with by nature of how it works.

It builds its answer from scratch every time. That is fine for one request. It means the engine cannot reuse work between requests. Every route pays the full cost of scattering stepping stones again.

Shortcuts have to be found and added one port at a time, by hand. We have found and verified eight so far. There are more ports out there with the same kind of hidden bay or channel that we have not gotten to yet.

It only thinks about distance. It cannot yet answer questions like which path avoids bad weather this week, or which path is fastest right now given the currents, because it has no way to represent those things at all.

Route Engine 2.0 is being built to remove those ceilings. Not because today's engine is broken, but because we want a foundation that can eventually factor in weather, currents, and traffic, and that gets faster rather than slower as the map fills in over time. Think of it less as replacing something broken, and more as building the highway system after years of getting by on well-mapped back roads.

They are not competing. They are in sequence.

The Route Engine is live and running every route on the platform today. Route Engine 2.0 still needs to prove itself the same way the current engine did: real depth checks, real edge cases, real regression testing before it is trusted with a single real shipment.

Until then, the Route Engine keeps doing the job it is already good at, and 2.0 keeps getting built carefully behind the scenes. When it is ready, it will not replace the thinking that went into the first engine. It will be built on top of it.

The system suggests. The captain decides. Getting the suggestion right, whether it comes from the first engine or the second, is the whole point.