Breaking Barriers Hackathon

Real-time anomaly detection system deployed on AWS. Top 8 of 32 teams at AWS × Deloitte × AT&T.

Result: Finalist — Top 8 of 32 Teams  ·  Organizers: AWS × Deloitte × AT&T  ·  Stack: Python, XGBoost, AWS Lambda, S3, Amazon Location Service  ·  Year: 2025

GitHub →


Overview

Built a real-time crowd anomaly detection system for simulated tail-risk events (sudden crowd movements, stampede precursors) within a 24-hour hackathon. The system ingests LLM-generated synthetic event streams and surfaces anomalies on an interactive live map.

Data Pipeline

  • Implemented a 0.5–1s micro-batching pipeline to efficiently process high-velocity LLM-generated synthetic data streams in near real-time.
  • Engineered lagged spatio-temporal features via client-side sliding windows, enabling stateless backend processing with no session state required.

Model

  • Trained an XGBoost classifier for binary anomaly detection on a highly imbalanced dataset.
  • Tuned classification thresholds via PR AUC (Precision-Recall AUC) to aggressively minimize False Negatives, prioritizing recall for tail-risk events where missed detections carry high cost.

Deployment

  • Deployed ultra-low latency model inference via AWS Lambda with end-to-end response time <400ms.
  • Streamed anomaly triggers to an S3-hosted dashboard using Amazon Location Service and MapLibre for real-time geospatial visualization.
  • Architecture is fully serverless and auto-scales with event volume.