Building Oi! at AWS Summit Taipei 2026
ICTALAB turned 21,282 vessel Noon Reports into Oi!, an AI powered AWS platform for vessel energy efficiency and maintenance decision support.
At AWS Summit Taipei 2026, my teammates from ICTALAB at National Chung Hsing University and I joined the three day AI Everywhere Hackathon and took on a challenge proposed by Yang Ming Marine Transport Corporation.
The challenge focused on vessel performance and energy efficiency.
We quickly realized that predicting fuel consumption alone would not be enough. Fleet managers also need to know which vessel needs attention, what may be causing the performance change, and whether maintenance should be considered.
That became the idea behind our project:
Oi! Vessel Energy Efficiency Operations Decision Platform
Instead of building only a machine learning model, we wanted to connect vessel data, prediction, explainable AI, maintenance planning, and an AI advisor into one decision workflow.
After three intensive days, our project won first place in the Yang Ming Marine Transport challenge.
From fleet monitoring to action
We designed Oi! around how fleet managers might actually use the system.
The platform starts with a fleet dashboard showing the status of 15 vessels, maintenance schedules, Speed Loss trends, and alerts. Users can then select an individual vessel to review its operational indicators, historical Noon Reports, and related events.

Figure 1. Oi! provides a fleet wide view of vessel status, Speed Loss, maintenance activities, and operational indicators.
The idea was simple.
Before asking users to understand a machine learning model, first help them answer:
Which vessel should I look at?
See Oi! in action
We also built a working prototype during the hackathon.
Oi! Demo
The demo shows the flow from fleet monitoring and vessel analysis to maintenance decisions, alerts, and the AI fleet operations advisor.
For us, this was important. We did not want to finish with only a model and evaluation metrics. We wanted to turn the models into something people could actually use.
Starting with 21,282 Noon Reports
Our work started with 21,282 Noon Reports.
After removing duplicates and filtering the data based on sailing hours, wind conditions, valid fuel consumption, and target availability, we prepared 8,057 records for formal model training.
We also normalized different fuel types into a common VLSFO equivalent 24 hour fuel consumption measure so they could be modeled on the same scale.
The final feature set contained 62 features, including vessel speed, RPM, propulsion related variables, nonlinear features, and additional features for low speed conditions.
The model was therefore learning more than just the relationship between speed and fuel consumption.
We wanted it to also consider how much propulsion effort was required to produce the vessel's actual movement.
Prediction is only useful when it leads to a decision
We compared several modeling approaches and selected CatBoost direct prediction as our final approach. It achieved a validation MAE of 2.3901 MT, a WAPE of 3.34%, and an R² of 0.9836.
We also used SHAP to understand how the model was making its predictions.
RPM related features were among the strongest signals, followed by vessel speed and propulsion related variables. We treated SHAP as model explanation rather than physical causality, which is especially important when the output may influence maintenance decisions.
But prediction was still only the first step.
Oi! can also simulate different maintenance actions, such as hull cleaning and propeller polishing, estimate future trends, and suggest possible maintenance timing and priorities.
We also brought market fuel prices into the workflow so the system could move from a technical prediction toward a business decision.
The question we ultimately wanted to answer was not simply:
How much fuel will this vessel consume?
It was:
When does it make sense to take action?
Building Oi! on AWS
AWS was at the center of the project, but the three day time limit also influenced our architecture decisions.

Figure 2. AWS architecture of the Oi! hackathon prototype.
For the hackathon prototype, we deployed the web application on Amazon EC2.
This helped us move quickly, integrate different components, and troubleshoot the system within the limited development time.
However, this would not be our preferred architecture for a production deployment.
With more time, we would move toward a serverless design.
The static frontend could be hosted on Amazon S3, backend application logic could run on AWS Lambda, and operational application data could be stored in Amazon DynamoDB.
This would reduce infrastructure management and allow different parts of the application to scale independently.
So the EC2 deployment was a practical hackathon decision, not our final production architecture.
I think this is also an important part of architecture design.
The best architecture depends on the constraints.
For a hackathon, development speed matters. For production, scalability, maintainability, cost, and operational overhead matter much more.
An AI fleet operations advisor
Oi! also includes an AI fleet operations advisor.
Users can select a vessel and ask questions about its operational information or related knowledge. The advisor returns answers together with relevant information sources so users can understand where the response came from.
We do not see AI as a replacement for fleet managers.
Its role is to help organize large amounts of data, explain what is happening, and make useful information easier to reach.
The final decision still belongs to the person responsible for the vessel.
Why an LLM Wiki?
We also explored an idea we called an LLM Wiki.
RAG is useful for retrieving information for the current question, but we were also interested in what happens after a useful question has been answered.
Can that knowledge be kept and reused?
Our LLM Wiki stores structured knowledge, sources, versions, maintenance experience, and previous analysis results. An agent can search this knowledge, reason across it, and write useful results back into the Wiki.
The idea can be summarized as:
Ask → Analyze → Answer → Update Knowledge → Reuse
We see it as a complement to RAG.
RAG helps answer the current question.
The LLM Wiki is intended to help useful knowledge remain available for future questions.
What comes next
The current model is mainly based on daily Noon Reports.
With more detailed data such as GPS, AIS, weather, ocean currents, shaft power, torque, real time fuel flow, propeller slip, and underwater inspection records, we could improve both prediction accuracy and the ability to explain vessel performance changes.
Adding fuel prices, carbon costs, cleaning costs, vessel schedules, and maintenance constraints could move the platform even further toward real operational optimization.
The direction is clear:
From predicting fuel consumption to supporting better operational decisions.
Three days later
A three day hackathon forces you to make choices quickly.
We had to decide what to build, which models to test, what architecture we could realistically finish, and how to connect the technical work to actual business value.
By the end, Oi! had become our attempt to connect:
Data → Prediction → Explanation → Decision → Knowledge

Figure 3. ICTALAB receiving first place in the Yang Ming Marine Transport challenge at the AI Everywhere Hackathon during AWS Summit Taipei 2026.
I am very happy that our project received first place, and I am grateful to Professor Yi Chung Chen and my ICTALAB teammates for working through these three intensive days together.
For me, the biggest takeaway was simple.
A useful AI system is not just about building a good model.
It is about connecting the model with data, cloud architecture, domain knowledge, and the decision someone eventually needs to make.
That is what we tried to do with Oi!.
And for three days, I think it was a pretty good start.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article