Research
Peer-Reviewed Publications
Detection systems for data that cannot be trusted at the source.
-
Secure and transparent energy management using blockchain and machine learning anomaly detection: A case study of the Ausgrid dataset
Decentralized energy data architecture combining blockchain immutability for smart-meter audit trails with unsupervised machine learning pipelines to detect consumption anomalies.
- Data pipeline & auditability: ingests high-frequency smart-meter time-series into a decentralized ledger to ensure tamper-proof consumption records.
- Unsupervised anomaly engine: benchmarks data restoration and unsupervised anomaly annotation models over smart-meter streams to catch power theft and sensor anomalies.
- Empirical validation: benchmarked against 300 residential solar and grid consumption feeds from Australia's Ausgrid smart-meter dataset.
-
Dual verification-based multimodal approach for reliable pothole and speed bump detection on embedded system
as of August 2026
Edge-computing data pipeline combining accelerometer telemetry and vision streams with dual-verification consensus, preventing single noisy sensor channels from triggering false hazard alerts.
- Multimodal sensor stream fusion: synchronizes IMU motion telemetry with visual detection models on resource-constrained embedded targets.
- Dual-verification consensus: requires agreement between inertial and vision channels before registering a road anomaly, eliminating false positives from single-sensor noise.
Technical Articles
The explanation I wanted when I was building it.
-
Simplify Your Log Management: Configuring DataDog Integration with Log4j2 In Mulesoft
- A Log4j2 HTTP appender POSTs JSONLayout events straight to Datadog's Send Logs API.
- No agent on the host, and no collector in between.
- The appender is asynchronous, so logging never blocks a Mule execution thread.
-
Developing a Custom YOLOv8 Model for Number Detection on Meters Using FastAPI and Gradio
- Annotates the meter dataset in Roboflow and trains YOLOv8 from it.
- Trains across Kaggle's two TPUs with DataParallel.
- Serves the weights twice: a containerized FastAPI endpoint, and a Gradio app on Hugging Face.