Raghunathareddy
GR

Founded and built an edge-native MLOps platform for deploying ML models to fleets of ARM devices — Jetson boards, Raspberry Pis — without brittle SSH scripts. A 7.4MB Go agent handles offline-first inference with on-device drift detection in under 2ms, pull-based delta model syncing, and one-command rollbacks. Shipped a 41-endpoint multi-tenant control-plane API, a published Python SDK and CLI (pip install mlops-dev), and a full marketing site with a waitlist and Discord community — built for regulated industries like medical devices and manufacturing that need an auditable record of exactly what model ran on which device, on which date.

Automated NLP classification pipeline with continuous integration, automated evaluation, and deployment. Evidently AI runs scheduled drift checks to catch data and target drift early, while the MLflow Model Registry manages versioning and promotion from staging to production. A GitHub Actions workflow runs linters and tests, builds Docker images, and auto-deploys to AWS EC2 on every change.

Containerized a computer vision model and orchestrated its deployment as a Kubernetes microservice. Horizontal Pod Autoscaling and tuned resource limits held inference latency under 100ms during traffic spikes, validated by load-testing 500+ concurrent users with Locust. Prometheus and Grafana dashboards track error rates, latency percentiles, and hardware utilization.

Built a data-centric pipeline for high-frequency transactional anomaly detection. Feast Feature Store serves consistent features across batch training and low-latency online inference, eliminating training-serving skew, while a 2GB dataset is version-controlled with DVC on AWS S3 for full pipeline reproducibility.