Ashwin Upadhyay

Spearheaded the design and implementation of an end-to-end, production-ready Suspicious Human Activity Recognition (SHAR) system utilizing a ViT-LSTM model for real-time video stream analysis. Developed a full-stack surveillance dashboard using Flask-SocketIO and React for low-latency, real-time model inference visualization and monitoring, showcasing full application deployment expertise. Engineered a robust MLOps pipeline leveraging Kubernetes (as per certification) and Apache Airflow (Certified DAG Author) for resilient, high-volume model training and serving. Trained the core model on a custom, large-scale dataset synthesized from UCF-Crime and KTH datasets, demonstrating advanced data preparation and ETL capabilities (Apache PySpark). Integrated Generative AI principles (RAG, Vector Search) into the project's architecture, applying expertise from Oracle AI Vector Search and Databricks Generative AI certifications to plan and prototype future semantic search capabilities within the surveillance logs. Architected the cloud-native infrastructure on Oracle Cloud Infrastructure (OCI), applying principles from multiple professional certifications (OCI Certified Architect and Developer), ensuring the system is secure, scalable, and optimized for high-throughput, event-driven inference workloads.

Automated CI/CD for model services with Git Hub Actions + Docker + Helm, reducing manual release steps by 30% and cutting deploy time from 45 to 20 minutes. Established a Kafka ingestion pipeline (12 Partitions, 20k messages/milliseconds throughput) delivering exactly once data into Snowflake via Snow pipe, reducing ingestion latency from 30 to 5 minutes. Hardened reliability with health checks, retry/back off, and alerting (Prometheus/Grafana), improving uptime to 99.5% for streaming components