Hi, I'm Junie Mariam Varghese 👋
Software Engineer · AI/ML Infrastructure · GPU Inference
I build reliable software and machine learning systems—from distributed services and cloud platforms to optimized model inference on GPUs.
Currently focused on:
- AI/ML infrastructure, model serving, and inference performance
- Kubernetes-native platforms, autoscaling, and distributed systems
- GPU computing with PyTorch, ONNX, TensorRT, and Triton
- Contributing to open-source projects in the cloud-native and AI ecosystem About me
- 🎓 M.S. Software Engineering, San Jose State University
- 💼 Former Software Development Engineer at Philips
- 🔬 Experience across backend engineering, cloud, DevOps, medical AI, and ML platforms
- 🚀 Interested in software engineering, ML platform, MLOps, GPU infrastructure, and systems roles
- 📍 Based in San Jose, California
Featured work
| Project | What I built | Results |
|---|---|---|
| GPU Cluster Reliability & Inference Control Plane | FastAPI control plane for GPU scheduling, health-aware routing, vLLM/Qwen serving, recovery, SSE streaming, Prometheus metrics, and Kubernetes deployment | 18 passing tests; benchmark tooling, admission limits, backpressure, and persistence validation |
| Genome2Expression | DNABERT and Enformer-lite pipeline with PyTorch, ONNX, TensorRT, Triton, Kubeflow, MLflow, and DVC | +17% accuracy; latency 420 ms → 130 ms; GPU memory −28% |
| AI Knowledge Transfer Assistant | Multimodal RAG and agent platform with FastAPI, LangChain, GKE, Kubernetes, and CUDA | 10,000+ documents; sub-300 ms latency; sub-minute synchronization |
| Smart AI Ops | ML operations platform with scikit-learn, XGBoost, MLflow, FastAPI, Docker, and Kubernetes | PR-AUC 0.724 → 0.929; ROC-AUC 0.874 → 0.976; F1 0.77 → 0.889 |
| Customer Intelligence & Experimentation Platform | End-to-end customer analytics pipeline with PySpark, scikit-learn, XGBoost, MLflow, SHAP, and FastAPI | Churn ROC-AUC 0.7572; value-model R² 0.9004; simulated campaign lift +3.70 pp |
| SafeSort-RL | PPO-based robotic sorting policy in Genesis using Stable-Baselines3 and AMD ROCm/HIP | 81% balanced success across 100 evaluation runs |
| MiniDFS | HDFS-inspired distributed filesystem in C++ with a metadata master, replicated storage nodes, consistent hashing, multithreaded TCP servers, integration tests, and Valgrind-based concurrency debugging | Fault-tolerant three-node replication with concurrency and performance testing |
Open source I contribute to cloud-native and AI infrastructure projects, including:
- Kubeflow — Kubeflow community — validation improvements and dependency/image maintenance
- Numaflow — Numaproj — OTLP metrics export for Daemon services
- KAI Scheduler — NVIDIA — autoscaling and scheduling improvements
- Grove — NVIDIA Dynamo — native KEDA integration for AI workload autoscaling
- Argo Helm — Argo Project — Helm chart work for the Argo CD MCP Server
Tech stack
- Languages - Python · C++ · Go · TypeScript · JavaScript · SQL
- AI/ML and inference - PyTorch · Hugging Face · ONNX · TensorRT · Triton Inference Server · vLLM · CUDA
- Cloud and platforms - Kubernetes · Docker · AWS · GCP · Azure · FastAPI · PostgreSQL · Neo4j · ChromaDB
- MLOps and developer tools- Kubeflow · MLflow · DVC · Terraform · Jenkins · GitHub Actions · Linux · Git
Connect Feel free to explore my repositories or connect with me about software engineering, AI infrastructure, and open-source systems.