1

Enabling Local AI for Federal Maritime Intelligence

Customer Case Study

Federal agencies processing sensitive imagery face a fundamental technology gap: commercial laptops lack the computational power to run capable AI models locally, while cloud-based solutions introduce unacceptable data security risks for classified or sensitive operations.

A Vision-Language Model Workflow on the HP ZGX Nano Curtis Burkhalter, Ph.D., HP AI Solutions Product Manager

Executive Summary This case study demonstrates a practical AI workflow for maritime surveillance intelligence generation using the HP ZGX Nano G1n AI Station, proving that local AI development with production-grade models is both achievable and cost-effective for regulated environments.

2

The real constraint is not interest in AI; it is infrastructure. Agencies need the ability to run advanced models directly inside secure environments.

This allows agencies to bring advanced AI capabilities directly to mission environments, enabling secure model execution where sensitive data already resides.

Demonstration: Maritime Surveillance Intelligence Generator To illustrate the practical value of local AI development, I built a Maritime Surveillance Intelligence Generator that processes aerial reconnaissance imagery and produces structured threat assessment reports. This workflow demonstrates capabilities directly relevant to Navy operations: automated vessel classification, cargo identification, activity assessment, and threat level determination.

AI developers working with sensitive data face a difficult choice. Standard commercial hardware, even high-end laptops, cannot run large language models and vision-language models that deliver production-quality results. A laptop with 16-32GB of RAM simply cannot load a 32B or 70B parameter model that would provide the accuracy needed for critical applications.

Compounding the challenge, sending data to cloud AI services is often not an option. Federal agencies handling reconnaissance imagery, intelligence reports, or other sensitive materials cannot transmit this data to third-party cloud providers. This creates a capability gap where organizations are forced to use smaller, less capable models that may not meet operational requirements.

The real constraint is not interest in AI; it is infrastructure. Agencies need the ability to run advanced models directly inside secure environments they control, without exposing sensitive data or sacrificing performance.

The HP ZGX Nano G1n AI Station addresses this gap directly. With 128GB of unified LPDDR5x memory and NVIDIA’s Grace Blackwell architecture, it enables local inference of models up to 200B parameters and fine- tuning of models up to 70B parameters, all in a form factor smaller than a standard desktop PC.

This allows agencies to bring advanced AI capabilities directly to mission environments, enabling secure model execution where sensitive data already resides.

The Challenge: AI Capability vs. Data Security

3

Figure 1: Maritime Surveillance Intelligence Generator output showing vessel classification, threat assessment, and actionable recommendations.

This demonstration validates two critical capabilities for federal procurement decision-makers:

Data sovereignty without capability compromise. Sensitive imagery never leaves the local device. Agencies can work with proprietary or classified data using state-of-the-art open-source AI models without relying on cloud infrastructure or third-party services. The models used in this demonstration, BLIP-2 and TinyLlama, are fully open-source.

Value Proposition for Federal Agencies

Threat levels are determined programmatically based on vessel classification, with military vessels flagged as HIGH priority.

Foreign national vessels flagged for increased monitoring, and commercial vessels categorized appropriately based on cargo type and activity.

Architecture The system combines two open-source models running entirely on the ZGX Nano. Salesforce/BLIP-2 FLAN-T5-XL serves as the vision-language model, performing visual question answering to extract structured information from imagery. The model answers targeted questions about vessel type, cargo, activity status, size, and heading. TinyLlama/TinyLlama-1.1B- Chat-v1.0 enhances the extracted information into natural language threat justifications and contextual observations.

The workflow processes images through a series of maritime-specific questions.

4

The HP ZGX Nano G1n AI Station eliminates the forced choice between AI capability and data security. Federal agencies no longer need to compromise on model performance due to hardware limitations or risk sensitive data exposure through cloud services. This maritime surveillance demonstration proves that production-grade vision-language model workflows can run entirely on local hardware, providing the computational foundation for secure, capable AI development in regulated environments.

Path to Production. This demonstration used general-purpose open-source models without domain-specific fine-tuning. Organizations interested in deploying similar capabilities could significantly improve accuracy by fine- tuning models on their specific vessel recognition requirements, regional imagery characteristics, or classification taxonomies. The HP ZGX Nano supports fine-tuning models up to 70B and 200B parameters at FP4 precision, enabling agencies to develop and use custom models tailored to their operational needs while maintaining complete data control throughout the training process.

Reference technical materials supporting this demonstration are available for further exploration.

Solving the Local AI Challenge for Classified Workloads

Edge deployment readiness. The ZGX Nano’s compact form factor (5.9” x 5.9” x 2.1”) and 240W power envelope makes it suitable for deployment in forward operating environments, ships, or mobile command centers where low-latency inference is required and network connectivity to cloud services may be unavailable or insecure.

Learn more about the HP ZGX Nano and see it in action — request a demo

© Copyright 2026 HP Development Company, L.P. The information contained herein is subject to change without notice. The only warranties for HP products and services are set forth in the express warranty statements accompanying such products and services. Nothing herein should be construed as constituting an additional warranty. HP shall not be liable for technical or editorial errors or omissions contained herein.