FraudShield AI

graph representing anomaly score
showing international transactions

Inspiration

FraudShield AI was inspired by the growing challenge of detecting financial fraud in real time. Traditional rule-based systems often miss subtle patterns or require manual monitoring. I wanted to explore how AI, combined with vector embeddings, could enhance fraud detection by identifying anomalies that are contextual and semantic, not just numeric.

What it does

FraudShield AI detects suspicious financial transactions by combining:

Vector embeddings of transaction descriptions for semantic search.
Rule-based checks for amounts, keywords, and patterns.
Anomaly detection using Isolation Forest to flag outliers. It then assigns a risk severity (Low, Medium, Suspicious, High) for each transaction, helping analysts prioritize investigations.

How we built it

Used Python FastAPI to create a backend API.
Integrated TiDB Cloud for scalable SQL storage and vector search.
Embedded transaction descriptions with sentence-transformers (384-dimensional vectors).
Applied rule-based logic and Isolation Forest for anomaly scoring.
Connected everything securely using SSL certificates and environment variables (.env) for secrets.

Challenges we ran into

Configuring secure connections to TiDB Cloud and handling SSL certificates.
Efficiently updating embeddings in batches without overloading the database.
Balancing ML-based anomaly detection with traditional rule-based flags to reduce false positives.

Accomplishments that we're proud of

Built an end-to-end AI + rule-based fraud detection system that runs on TiDB Cloud.
Successfully integrated vector embeddings with SQL queries, enabling semantic search on transactions.
Automated scoring and severity assignment, reducing manual review effort.

What we learned

How to implement vector embeddings and similarity search for real-world data.
Best practices for secure, cloud-based database access.
How to combine ML models with rule-based systems for a robust fraud detection pipeline.