From Diesel to Data: Logistics Prediction and Text-to-SQL Pipelines with AWS AI

Discover how a five-week Generative AI sprint transformed manual post-trade analysis into a predictive powerhouse, achieving 79% accuracy in logistics forecasting and delivering a secure, natural-language chatbot for instant data insights.

ClientPost-Trade Logistics

PublishedFebruary 2026

IndustryEnergy & Commodities

Most trade operators spend their days digging through historical data just to guess when a shipment might arrive. This case study breaks down a 5-week project that replaced those manual workflows with a prediction engine hitting 79% accuracy and a secure chatbot that lets users query complex trade data in plain English.

Our customer is a technology company, founded by a consortium of major industry players, that delivers a sophisticated post-trade processing platform for the global commodities industry. The company's core mission is to digitize and automate the complex, contract-heavy, and logistically intensive post-trade lifecycle. Their platform focuses on digitizing trade contracts and managing the complex global logistics required for the movement of physical commodities.

The customer's platform processes vast amounts of sensitive trade and logistics data, but they faced several challenges in leveraging this data to its full potential. This project was initiated to validate the feasibility of new, intelligent features while adhering to their extremely high standards for data security.

Manual & Repetitive Data Analysis: The customer's users, such as trade operators, performed repetitive analysis of historical data to predict logistics actions, like the timing and quantity of future nominations. This process was manual and time-intensive.
Inaccessible Data Insights: Business managers and operators could not easily query the rich historical trade data. Accessing specific insights, like "when was the last time asset X was used," required custom-built reports, which was a slow and inflexible process.
Complex and Nuanced Data: The customer's existing systems struggled to accurately interpret and process highly complex documents and nuanced pricing structures, often omitting critical information.
Strict Data Privacy & Isolation: The most significant challenge was the platform's federated architecture. As a system for competing global companies, each customer's data is strictly isolated in its own dedicated instance. Any AI solution must respect these boundaries and guarantee that data from one client could never be seen by another.

A serverless AWS architecture was designed and deployed, delivering two distinct Generative AI solutions running in the customer's isolated AWS account.

Secure Application & API Layer
To demonstrate the solution, a functional demo application was built.

Amazon API Gateway: Amazon API Gateway was used to create both an HTTP API (for predictions) and a WebSocket API (for the chatbot).
AWS Fargate & CloudFront: A demo UI built with Streamlit was hosted using AWS Fargate and an Application Load Balancer, with Amazon CloudFront providing secure, low-latency access to the application for the customer's team.

The Logistics Prediction Engine

To solve the challenge of manual forecasting, Stormit built a machine learning pipeline to predict three key outcomes for a new trade: the logistics pattern, the nomination timing, and the parcel quantity.

Amazon SageMaker was used to train an XGBoost model on the historical, anonymized data. The final model was deployed on a SageMaker Serverless Inference endpoint, providing a scalable, cost-effective way to get real-time predictions without managing servers.
An AWS Lambda function was created to process incoming prediction requests. It takes data from a new trade, calls the SageMaker endpoint for a prediction (model inference), and returns the result.

The Conversational AI Chatbot
To make data insights accessible, Stormit built an AI-powered chatbot that allows users to ask questions in plain English. This solution respects the platform's data isolation rules by not training a central model on the data.

Amazon Bedrock: The solution uses Amazon Bedrock to provide the Claude foundation model.
Intelligent Querying (Text-to-SQL): When a user asks a question, a "Chatbot Lambda" function securely sends the question and the database schema to Amazon Bedrock. Bedrock translates the natural language question into a precise SQL query.
Amazon Athena: The generated SQL query is then run against Amazon Athena, a serverless query engine. Athena queries the data directly in the Amazon S3 data lake, which contained the anonymized trade and logistics data in Parquet format
AWS Glue: An AWS Glue Catalog was used to define the data schemas, making the S3 data easily queryable by Athena.
The query results are sent back to Bedrock to be translated into a natural language answer for the user.

Both the Logistics Prediction Engine and the Conversational AI Chatbot were successfully delivered.
The prediction models yielded highly promising and measurable results:

Logistics Pattern Prediction: The model achieved 79% accuracy in predicting the correct logistics pattern.
Nomination Timing Prediction: The model predicted the timing of nominations with an average error of just 0.98 days.
Parcel Quantity Prediction: The model effectively explained 62% of the variability in the first parcel quantity data.¨

By leveraging a serverless AWS architecture with Amazon SageMaker and Amazon Bedrock, the customer was provided with a powerful, secure, and scalable foundation. The project gave the customer the concrete data and confidence needed to define a roadmap for integrating these intelligent features into their core platform, ultimately enabling them to offer unparalleled, data-driven value to their clients.