A Look into the “Engine Room” of Windward’s AI

Data is data, right? Is there really a difference between selecting Windward’s more than a decade of historical maritime data, or just buying automatic identification system (AIS) data? It turns out there is a big difference. 

First, you would need to spend millions of dollars to obtain ten years of historical AIS data – and working with less than that puts you at risk of not having sufficient data to optimally train and refine AI data models in the complex maritime ecosystem. 

Additionally, we should consider the saying, “Garbage in, garbage out.” AIS data comes with a great deal of “noise,” and effective AI is not just about data, it involves taking the data and producing actionable insights.

Windward’s Data Differentiators

• Multiple sources – by utilizing multiple sources for data, Windward ensures our technology is not over-reliant on any one source. 
• Clean data – some maritime technology vendors show data as-is, while others do basic clean-up. Windward heavily invests in cleaning the data, because if data is not clean, everything built on top of that flawed foundation will be wobbly. Maritime domain expertise is key for understanding and constantly evaluating data points.
• Iterative improvement of current and historical data – part of what sets Windward apart is our platform’s ability to take new insights and apply them to our historical data. This transforms the existing data into a treasure trove that continues to yield new analytical gems.

Windward underwent an important upgrade for our customers by migrating data into Rockset, the real-time analytics database built for the cloud. We evaluated ten databases to find the one that would best support our ability to serve our customers.  

Our complex use cases and huge amount of data require an analytical database that offers top performance and we are already seeing the benefits. Extremely complex queries that were not previously possible are now being executed in a few seconds or less!

Analyzing Windward’s Analytical Database

• Everything is searchable. This makes it much easier to query new data.
• On demand queries instead of “pre-baked” results – for instance, users can create  “custom polygons” to monitor specific maritime areas of interest. Using the new database, Windward’s Maritime AI™ can run queries on historical data for newly-defined polygons and produce results in seconds, to support ad hoc investigations of areas of interest. 
• Faster response time – this spans from development, all the way to running a query and receiving the results.
• Allocate dedicated resources for customers – users can request an increased number of APIs compared to our default, via a dedicated server that handles all queries. Users expecting a large increase in expected queries for whatever reason can now go beyond the API limitations.

You want to watch a video or TV show, so you open Netflix, select some content, and it immediately begins playing. But think back even a few years, that experience was considerably slower. You would click “Play” and then…inevitably wait for the buffering. 

Earlier iterations of AI for the maritime ecosystem worked in batch mode, which meant buffering was part of the process. Systems collected a great deal of data and then analyzed it all together. This took time and once the analysis was finished, it was shared with users. 

The streaming method now used by Windward means that once a new datapoint is received, instead of waiting for an unwieldy process to occur, our platform takes only that new datapoint and analyzes it (along with all the dependencies). If we get a new blip, Maritime AI™can calculate and then add it to the relevant vessel. If there is risk that is dependent on that blip, it will also be updated.

Streaming greatly reduces latency, which is important, because many incidents are time-sensitive. To give an example, certain government agencies must catch illicit actors who have entered the polygon immediately, before they flee. 

Another example, which shows how Windard’s AI-powered technology saves organizations time and money, while also enabling more business, is vessel selection. 

A maritime organization wants to understand if they should charter a specific vessel for a journey. By providing real-time risk analysis of that vessel, Windward helps them avoid a vessel they shouldn’t use – because it is not fuel efficient, or for whatever other reason. Avoiding the extra expenditure would have been impossible if not for the reduced latency. The flip side of the coin: if a vessel was blocked and then Windward’s AI quickly unblocked it, users do not lose a day before being able to engage in new trades.

REST APIs are extremely familiar to most technical personnel, but Windward has selected GraphQL technology — why? 

GraphQL makes it simpler for users to request what they want. Once we understood that customers wanted our APIs, we just had to figure out the best way to expose them and offer greater flexibility, instead of developing ad hoc solutions for every client. This significantly shortened time to market. It depends on data complexity, but integration can happen in just a few hours.

With GraphQL, the user can specify everything he or she wants, without multiple requests. This is a big difference compared to REST, which limits the data you can receive and the format you will receive it in. GraphQL can save IT/development teams days, or even weeks, by providing the flexibility to do things their way.

Fast integration and user agility are critical in today’s world, as large-scale events keep transforming the maritime industry. First we experienced the largest global pandemic of our lifetimes and right as that seemed to be abating, Russia unexpectedly decided to invade Ukraine. Major climate events are also becoming increasingly likely. IT teams must be equipped to proceed quickly. 

If you build a REST API and the world transforms, you have to totally rebuild, which will be time consuming and costly. GraphQL is easy to build, so you do not have to waste six figures every time. 

With GraphQL, there is a simplicity that empowers users to “play” with an API. You develop an API, it is automatically documented, and you can experiment with it. The APIs are strongly typed, which prevents errors by showing type details and blocking incorrect type choices.

Windward’s API Advantages

• Extreme flexibility – as noted above, users can quickly add more fields and information. 
• Automatic notifications via webhooks – they can change the functioning of a webpage or web application with custom callbacks, and webhooks are compatible with GraphQL. These automated answers are extremely convenient for users. They are able to create a prompt to ask whether a vessel arrived in a polygon, for instance, and then users can rest assured knowing that no further action is required to receive the eventual notification.

It often takes years to develop and deploy even simple AI models and systems. As noted in the previous section of this Executive Brief, this can be problematic due to the sudden and major shifts the world sometimes experiences. But we do not need to focus solely on black swan events. Changes in political regimes, new regulations or sanctions, a shift in trade routes, etc., can necessitate quick change. 

A failure to adapt in an evolving world does not mean staying in the same place, but rather sliding backwards. AI models quickly become less and less relevant.

A key to crossing the AI development chasm is reducing time to market via effective machine learning operations (MLOps). Understanding that requires first understanding the different MLOps levels. 

3 MLOps Levels

Level 0 indicates that the development team is managing less than ten AI models manually. Most companies that are utilizing AI are currently at this level. 

Level 1 means automation of the deployment process of a model – this is Windward’s current status. We can deploy models relatively easily. For instance, creating our Ocean Freight Visibility solution involved expediting the development of two new models: Vessel ETA and Container ETA. 

Upgrading to level 1 meant that time to release went from several weeks, to minutes! This was obviously a significant acceleration of the time to take a model from research to production.
Organizations at level 1 manage more than ten models. 

Level 2 includes continuous integration and development. It empowers data scientists to rapidly explore new ideas around feature engineering, model architecture, and hyper-parameters. At this stage, organizations rapidly and automatically retrain models and can get to more than 1,000 AI models. Windward has already made several strong steps to eventually ascend to this level.

MLOps Opportunities

As noted, it is important to release new models often in a changing world. Once you have an AI model that was released to production, you must keep training it. Windward releases new models every other month. 

Every time a new version is released, results of the new model get compared to Windward’s established baseline – and of course we only release the model if it performs better. Finishing development of the model is merely the starting point, then comes the training. 

In addition to releasing new models, retraining existing models and continuously monitoring them are critical. 

Another functionality Windward is working on is being able to create a data structure that holds data in a way that is optimized for research. A data stream that is primed for research and enriched with customer data will be a game-changer in terms of producing actionable insights. 

AIS and operational data are difficult to use for research purposes, especially due to the “noise” of AIS data. Windward’s new data structure solves this issue.