A config-driven orchestration platform for deploying distributed agentic AI systems, cutting deployment to a single zip upload or one docker compose up.

• Provisions apps across 15 Docker-in-Docker worker nodes via a FastAPI control API, a Kafka message bus and a CLI (init, validate, build, package, deploy)
• Scripted node provisioning (SSH key generation, image build, network attachment) and orchestrated the Control API, Kafka and Zookeeper in one Docker Compose stack

A real-time computer vision pipeline for traffic-signal detection and counting from dashcam footage.

• Fine-tuned a pretrained YOLOv8 model via transfer learning on a custom-annotated dataset
• Augmentation pipeline (brightness/contrast jitter, motion blur, synthetic fog/rain) and IoU-based NMS tuning to cut false positives
• Geotagged detections with GPS metadata and used DBSCAN clustering to deduplicate repeated sightings of the same signal

A novel Byzantine fault-tolerant clock synchronization protocol combining Brooks-Iyengar sensor fusion with Inter-Tertile Range (ITR) outlier detection to achieve bounded-error time agreement in adversarial peer-to-peer networks, tolerating up to N/3−1 malicious nodes.

• Switched from quartile- to tertile-based outlier filtering, raising Byzantine tolerance from 25% to the theoretical 33% bound
• Applied Brooks-Iyengar interval fusion to aggregate uncertainty ranges into a single consensus offset with a quantified confidence interval