MarineGuard

• 

• 

• 

Inspiration

Offshore marine operations depend on engineering evidence.

For a Marine Warranty Surveyor (MWS), determining whether a complex operation is ready to proceed can require reviewing lifting calculations, installation procedures, vessel information, metocean criteria, subsea documentation, risk assessments, drawings, certificates, and other project records.

The challenge is not simply finding information. The real challenge is connecting the information:

What evidence supports which requirement?

Where did the evidence come from?

Has it been verified?

What is still missing?

Are there conflicts or gaps?

MarineGuard was inspired by this practical problem. We wanted to build an AI-powered assurance system that could reduce the manual effort involved in engineering evidence review while preserving traceability and keeping the final professional decision with the Marine Warranty Surveyor.

---

What it does

MarineGuard is an AI-powered Marine Warranty Survey assurance engine designed to go beyond a simple document chatbot.

When an engineering document is uploaded, MarineGuard resolves the project and operation context, extracts engineering evidence, links that evidence to its source, maps it against assurance requirements, and presents the resulting readiness state for MWS review.

From a 37-page calculation to structured engineering evidence

For our demonstration, MarineGuard processes a 37-page Check of Lifting Arrangement - Calculation Note for an EGINA subsea jumper installation workflow.

MarineGuard in Action

Within 45 seconds, MarineGuard extracted 42 engineering evidence parameters.

Examples include:

Engineering Evidence	Extracted Value	Source / Traceability
Jumper weight in air	22,860 kg	Check of Lifting Arrangement - Calculation Note, Section 4.3, Page 11
Jumper weight in water	15,935 kg	Check of Lifting Arrangement - Calculation Note, Section 4.3, Page 11
Developed jumper length	47.45 m	Check of Lifting Arrangement - Calculation Note, Section 4.1, Page 7
Pipe dimensions	OD 168.3 mm / WT 18.3 mm	Check of Lifting Arrangement - Calculation Note, Section 4.1, Page 7
Design pressure	517 bar	REF 1: DU600105288 Rev A, PJPN1C EGINA jumper drawing

For example:

jumper_weight_air

22,860 kg
Source: Section 4.3, Page 11
Extraction confidence: 85%

MarineGuard does not simply return a number. It preserves the relationship between the extracted engineering parameter and its source.

From extraction to assurance action

MarineGuard does not stop at answering questions about an engineering document.

Once a document is uploaded, it:

1. Resolves the operation context.
2. Extracts structured engineering evidence.
3. Preserves source-level traceability.
4. Determines the applicable assurance requirements.
5. Places evidence into the MWS verification workflow.
6. Maps evidence to requirements.
7. Surfaces outstanding evidence and assurance actions.

The result is a workflow that moves from document ingestion to structured assurance action, rather than simply generating a conversational response.

Human-in-the-loop verification

MarineGuard deliberately separates AI extraction from engineering verification.

The evidence workflow distinguishes between:

Evidence Status	Count
Extracted	42
Awaiting MWS Verification	40
Verified	2

The MWS can inspect the extracted value and source before explicitly verifying the evidence.

AI assists the assurance workflow. The qualified MWS retains the professional assurance decision.

Requirement intelligence

MarineGuard maps extracted evidence against the requirements applicable to the selected operation.

For the demonstrated Subsea Installation workflow, the system evaluates 15 assurance requirements.

The result is an assurance ledger rather than a simple document summary.

The MWS can see which requirements have supporting evidence, which require verification, and which still have no evidence.

Assurance readiness

MarineGuard turns the evidence and requirement status into a transparent readiness picture.

Instead of hiding uncertainty, the system surfaces:

• Evidence requiring MWS verification
• Requirements awaiting evidence
• Verified evidence
• Requirement compliance status
• Conflicts and gaps
• Outstanding assurance actions

The final assurance decision remains with the qualified MWS.

---

How we built it

MarineGuard was built as a modular application that separates the AI workflow from the underlying assurance logic.

The workflow connects document ingestion, project and operation context, AI evidence extraction, evidence traceability, MWS verification, requirement mapping, and assurance readiness.

MarineGuard Assurance Workflow

Step	Assurance Stage	What MarineGuard Does
01	Engineering Document	Ingests engineering calculations, procedures, certificates, and operational documents.
02	Project / Operation Context	Identifies the relevant project, operation, vessel, location, phase, and document context.
03	Applicable Requirements	Determines the assurance requirements relevant to the selected operation.
04	AI Evidence Extraction	Extracts structured engineering parameters and technical evidence from the source documents.
05	Source Traceability	Links extracted evidence back to its document, section, page, and source citation.
06	MWS Verification	Places extracted evidence in the MWS verification queue for human review and certification.
07	Requirement Mapping	Maps verified and pending evidence against the applicable assurance requirements.
08	Assurance Readiness	Presents the overall readiness state, outstanding evidence, conflicts, gaps, and MWS actions.

Technology Stack

Layer	Technology
AI Model	Gemini 3.5 Flash
AI Platform	Vertex AI
Agent Framework	Google Agent Development Kit (ADK)
Backend	Python / FastAPI
Frontend	JavaScript / HTML / CSS
Persistent Storage	Firestore / Application Evidence Store
Deployment	Google Cloud Run

MarineGuard maintains project and operation context so evidence can be scoped to the correct operational workflow.

The application uses environment-based configuration for Google Cloud and Vertex AI settings rather than embedding credentials directly in the application.

Google Cloud deployment

MarineGuard is deployed on Google Cloud Run.

Deployment	Value
Service	marineguard
Revision	marineguard-00018-p9r
Region	us-central1
Traffic	100%

The AI workflow uses Gemini through Vertex AI, with the Google Agent Development Kit supporting the agentic workflow.

---

Challenges we ran into

One of our biggest challenges was maintaining context throughout the entire assurance workflow.

An engineering value by itself is not enough.

MarineGuard needs to maintain the relationship between every stage of the assurance process.

An engineering value is only useful when MarineGuard can connect it to:

• the source document
• the relevant project and operation
• the applicable requirement
• its verification status
• the resulting assurance state

This relationship is critical because the same engineering value can have very different implications depending on the operation, requirement, source document, and verification status.

Another challenge was ensuring that AI-generated extraction was never treated as automatic engineering approval.

We addressed this through explicit evidence states and a human-in-the-loop verification workflow.

We also needed to ensure that evidence was associated with the correct operation and evaluated against the appropriate operation-specific requirements.

AI output vs engineering assurance

A conventional AI system might return:

"The jumper weighs 22,860 kg."

MarineGuard needs to answer a much larger question:

What is the value, where did it come from, what requirement does it support, has it been verified, and what does it mean for the assurance state?

That distinction shaped much of the system architecture.

Another challenge was balancing automation with transparency. In engineering assurance, the system must be useful without hiding uncertainty or presenting AI output as professional engineering approval.

---

Accomplishments that we're proud of

We are proud that MarineGuard evolved from the idea of document analysis into an operation-aware assurance workflow.

Key results

Result	Demonstration
Engineering document	37-page lifting calculation
Evidence extracted	42 parameters
Extraction time	45 seconds
Assurance requirements	15
Automated tests	143 passed
Cloud deployment	Google Cloud Run
AI platform	Vertex AI
AI model	Gemini 3.5 Flash
Agent framework	Google ADK

We are particularly proud of the evidence traceability.

For example, the extracted value:

22,860 kg

can be traced directly to:

Check of Lifting Arrangement - Calculation Note → Section 4.3 → Page 11

The evidence is therefore not presented as an unexplained AI-generated value. MarineGuard preserves its connection to the original engineering document and its location within the source.

We are also proud that MarineGuard makes uncertainty visible rather than hiding it.

The system maintains an explicit evidence lifecycle:

Extracted → Awaiting MWS Verification → MWS Verified → Requirement Status → Assurance Readiness

In our demonstration, MarineGuard extracted 42 engineering evidence parameters, with 40 awaiting MWS verification and 2 already verified.

This creates a transparent assurance workflow where AI-generated evidence can be reviewed against its source before contributing to the overall assurance assessment.

AI assists the workflow. The MWS remains the professional decision authority.

Our latest full test run completed with:

143 tests passed

The deployed application was also validated through the live Google Cloud Run environment with a successful health check returning HTTP 200.

---

What we learned

The biggest lesson was that applying AI to engineering is not simply about extracting information.

The real challenge is maintaining:

• Context
• Traceability
• Verification status
• Requirement relationships
• Decision boundaries

A value such as 22,860 kg has limited meaning by itself.

Its value within an assurance workflow comes from knowing:

Which document?

Which operation?

Which requirement?

Where in the source?

Has it been verified?

What remains outstanding?

This led to one of the central principles behind MarineGuard:

AI should accelerate engineering assurance without hiding uncertainty or replacing professional judgement.

We also learned that agentic systems are most useful when they take meaningful actions across a workflow rather than simply generating a conversational response.

The most important distinction for MarineGuard is:

AI assists the assurance workflow.

The qualified MWS retains the professional assurance decision.

---

What's next for MarineGuard

The current system demonstrates the core assurance workflow using engineering documents.

The next stage is to expand MarineGuard into a broader continuous assurance platform.

Multi-document assurance

Process complete engineering and operational document packages rather than individual documents.

Cross-document evidence comparison

Compare evidence across calculations, procedures, drawings, certificates, and other project records.

Revision intelligence

Identify how document revisions affect previously extracted evidence and assurance requirements.

Evidence conflict detection

Automatically surface conflicting values or statements across documents.

Automated MWS reporting

Generate structured evidence registers and draft assurance reports from verified evidence.

Continuous assurance

Maintain the assurance state as new documents and revisions are uploaded throughout an offshore campaign.

Our long-term goal is to give Marine Warranty Surveyors an intelligent assurance assistant that reduces repetitive document work while preserving the traceability, transparency, and professional judgement required for offshore engineering assurance.

---

The principle behind MarineGuard

MarineGuard is not designed to replace the MWS.

It is designed to make the MWS's work more efficient, structured, and traceable.

MarineGuard turns engineering documents into traceable assurance evidence, while keeping the final engineering decision where it belongs: with the MWS.

Built With

• agents
• fastapi
• firestore
• gemini
• javascript
• python
• vertex