ClaimShield

AI detects hidden fraud networks in insurance claims and explains the evidence behind each risk score.

Comment

Inspiration

Insurance fraud costs the industry over $308 billion every year, but the most difficult cases are not isolated suspicious claims. They are coordinated fraud rings made up of connected actors like repeat claimants, fake medical providers, and unregistered repair shops that appear harmless when reviewed one by one.

Fraud investigators already think in networks of relationships, but most fraud detection systems still analyze claims individually.

This inspired us to ask a simple question:
What if fraud detection systems could think the same way investigators do?

That idea led us to build ClaimShield, an AI-powered fraud detection platform that uncovers hidden fraud networks, explains its reasoning, and helps investigators act faster.

What it does

ClaimShield helps insurance teams detect coordinated fraud rings in real time.

When an adjuster uploads a claim, ClaimShield automatically:

  • Builds a graph of entities in the claim (claimant, provider, repair shop) and links them through shared history
  • Runs a Graph Neural Network (GraphSAGE) to detect suspicious subgraphs that resemble organized fraud rings
  • Uses an Isolation Forest to detect anomalies such as inflated repair costs or suspicious claim timing
  • Generates clear evidence explanations so investigators understand why a claim is flagged
  • Automatically drafts an investigator-ready narrative report to speed up investigation workflows

Instead of just assigning a fraud score, ClaimShield reveals the relationships and signals behind the decision.

How we built it

Backend: FastAPI
Frontend: React + TypeScript

AI / ML Models

  • GraphSAGE Graph Neural Network for relationship-based fraud detection
  • Isolation Forest for claim-level anomaly detection

IBM Stack

  • watsonx.ai for AI-powered narrative generation
  • Db2 for structured claims storage
  • watsonx.governance for decision logging and AI audit trails

Security & Infrastructure

  • JWT authentication
  • Fernet encryption
  • Dockerized services

This architecture allowed us to combine network intelligence, anomaly detection, explainability, and governance into a single fraud detection workflow.

Challenges we ran into

Building ClaimShield required combining multiple complex systems into a smooth workflow.

Some of our biggest challenges included:

  • Integrating graph-based fraud detection, anomaly scoring, and explainability into a single pipeline
  • Keeping the analysis fast enough for a real-time demo while still demonstrating meaningful technical depth
  • Designing a system that works for a hackathon demo but could also scale toward real-world deployment
  • Ensuring decisions were explainable and audit-friendly for a regulated insurance environment

Accomplishments that we're proud of

  • Training and experimenting with our fraud detection models using Kaggle datasets to validate our ideas quickly
  • Designing a hybrid AI pipeline that combines GraphSAGE for relationship-based fraud detection with Isolation Forest for claim-level anomalies
  • Building an evidence-backed narrative layer that explains not just the fraud score but the signals and connections behind it
  • Creating a system that integrates AI detection, explainability, and investigator workflows into one product

What we learned

  • Fraud detection becomes much more powerful when you analyze relationships between actors, not just individual claims.
  • Explainability is critical in regulated industries like insurance where investigators must justify decisions.
  • Integrating enterprise tools like IBM watsonx.ai, Db2, and watsonx.governance helped us understand how real-world AI systems balance detection, transparency, and compliance.
  • Designing clean interfaces between machine learning, backend services, and the UI makes systems easier to prototype quickly and scale later.

What's next for ClaimShield

Next, we plan to expand ClaimShield with:

  • Investigator feedback loops to continuously improve detection accuracy
  • Larger fraud network datasets for stronger model training
  • Deeper IBM watsonx integrations for enhanced governance and explainability
  • More advanced network visualization tools to help investigators explore fraud rings interactively

Our long-term goal is to turn ClaimShield into a platform that helps investigators uncover complex fraud networks faster, more transparently, and with greater confidence.

Built With

Share this project:

Updates