BunkerGuard AI: Singapore's Intelligent Bunker Fuel Fraud Detection Platform

BunkerGuard AI is an intelligent fraud detection platform designed specifically for Singapore's bunkering industry. It adopts a four-agent LLM workflow architecture, integrating document OCR, risk scoring, and evidence report generation to provide a complete technical solution for maritime fraud detection. The project aims to address the issues of low efficiency, poor consistency, and difficulty in scaling with traditional manual review, realizing automated fraud detection and evidence generation through modern AI technologies.

Bunker fuel delivery is a core link in the maritime industry and also a high-fraud area. As the world's largest bunkering port, Singapore processes a large number of Bunker Delivery Notes (BDN) every year, and manual review can hardly identify potential risks effectively. Common fraud methods include quantity shortage, quality adulteration, document forgery, duplicate claims, etc. Traditional detection relies on manual experience, which has problems such as low efficiency, poor consistency, and difficulty in scaling. BunkerGuard AI is an automated system built specifically to address these pain points.

BunkerGuard AI adopts a four-agent LLM workflow architecture and combines cloud-native tech stack to achieve full-process automation. Core design principles: LLM is responsible for interpreting and organizing evidence, while fraud scoring, carbon value calculation, and approval rules remain deterministic.

Tech Stack Composition:

• Frontend: React18 + TypeScript + Vite6, integrated with Vercel AI SDK, Three.js, etc.;
• Backend: Python3.12 + Pydantic2, AWS Lambda serverless runtime, Amazon Bedrock providing Claude models;
• Data & Infrastructure: Supabase Postgres for operational data storage, Amazon S3 for file storage, AWS CloudFormation for infrastructure management.

Real BDN Document Ingestion

Supports upload of documents in multiple formats such as PDF and JPEG. The process includes validation, secure storage, OCR extraction, and structured parsing, extracting key fields like vessel name and IMO number and calculating confidence levels.

Four-Agent LLM Workflow

• Surveyor Agent: Normalizes BDN into a format processable by the system;
• Investigator Agent: Generates risk and anomaly packages;
• Compliance Agent: Identifies compliance issues such as missing licenses and signatures;
• Decision Agent: Outputs SIGN/REVIEW/REFUSE recommendations.

Anomaly Detection Engine

Detects various anomaly types including quantity mismatch, density anomalies, sulfur content violations, etc. Each anomaly contains information such as stable ID and severity.

Risk Scoring Engine

Calculates risk scores by combining factors like anomaly severity and supplier history, outputting scores, categories, decisions, and audit trails.

In addition, it has capabilities such as Exa intelligent enrichment (external context query) and carbon emission calculation (based on fuel quantity and emission factors).

S3 Storage Prefixes

• uploaded-bdn/: Original BDN files;
• ingestion-evidence/: Agent packages;
• evidence-packages/: Stage2 and 3 API outputs;
• generated-reports/: Evidence reports. The storage bucket adopts security measures such as AES-256 encryption and version control.

Evidence Report Content

Includes session ID, BDN comparison, anomaly findings, risk assessment, financial exposure, environmental impact, recommended actions, etc., with SHA-256 hash and persistent metadata attached.

1. User uploads BDN → validation, hashing, storage;
2. Lambda starts asynchronous workflow → UI obtains job ID;
3. OCR extracts fields → persistence;
4. Surveyor Agent normalizes → creates bunkering session;
5. Exa enriches → stores external context;
6. Investigator Agent calls Stage2/3 → generates anomalies and risks;
7. Compliance Agent evaluates → lists compliance issues;
8. Evidence package is written to S3 → audit package;
9. Decision Agent outputs recommendations;
10. Frontend polls → user views results.

BunkerGuard AI demonstrates the way to build intelligent and reliable vertical applications by combining LLM with deterministic systems. Core insights:

1. Clear human-AI collaboration boundary: LLM is responsible for interpretation, while deterministic systems are responsible for decision-making;
2. Evidence-driven design: Conclusions have traceable evidence chains;
3. Auditable architecture: Logs, hashes, and version control ensure compliance;
4. Domain knowledge encoding: Expert experience is transformed into rule engines;
5. Cloud-native practice: Serverless for elastic scaling. This project provides reference value for fields such as financial risk control and compliance review.