ML Platform & LLMOps Engineer — I build the infrastructure layer that makes LLMs work reliably in production.
- MS Data Science · University of San Diego · GPA 3.9 · Dec 2025
- Built 5 interconnected LLM infrastructure projects spanning inference observability, Kubernetes deployment, GPU profiling, benchmarking, and conversational AI ops
- 3 MCP servers published — Python MCP Weather Server · StoryForge Agent · Job Recommendation Engine
- Built AI Infrastructure Copilot — LangGraph + Qdrant RAG + Claude, reduces GPU incident investigation from 40min → 3min
- NLP researcher — BERT emotion detection, BART/T5 summarization, topic modeling with CI/CD
- 48 automated tests · CI/CD with Trivy scanning · Helm charts · OpenTelemetry tracing
- Sunnyvale, CA · Open to full-time · No sponsorship needed
Open to: ML Platform Engineer · LLMOps Engineer · AI Infrastructure Engineer · MLOps Engineer · Applied AI Engineer · Founding Engineer (AI)
Five repos. One coherent platform. Built to serve LLMs reliably at scale.
| Repo | What it does | Key tech |
|---|---|---|
| ai-inference-observability-platform | FastAPI proxy — TTFT, TBT, E2E in every response · 48 tests · CI/CD · ≤31ms P99 overhead | vLLM FastAPI Prometheus Grafana OpenTelemetry Helm K8s |
| KubeInfer | Production K8s inference platform · queue-depth HPA · GitOps · 12 alert rules · 1000 req/min | Kubernetes Helm NVIDIA GPU vLLM GitOps HPA PDB |
| KV-Cache-Profiler | Real-time KV cache hit rate · eviction detection · GPU memory pressure | vLLM Prometheus DCGM Docker |
| LLM-Inference-Benchmarking-Dashboard | Live TTFT · TPOT · ITL · E2EL charts · NVIDIA DCGM GPU metrics | Prometheus DCGM Docker vLLM |
| AI-Infrastructure-Copilot | Conversational GPU diagnosis · Helm config generation · 40min→3min investigation | LangGraph Qdrant RAG FastAPI Slack PostgreSQL |
Internet ──► nginx Ingress (TLS) ──► Request Router ──► vLLM Engine × N (NVIDIA GPU)
│ │
Queue-depth HPA PagedAttention
(vllm:num_requests_waiting) Continuous batching
│ │
Prometheus ──► Grafana Shared PVC
12 alert rules (model cache)
OpenTelemetry
Microsoft invented MCP · Copilot Studio, Claude Desktop, Cursor compatible
| Server | Tools exposed | Transport |
|---|---|---|
| Python-MCP-Weather-Server | check_weather(location) — structured JSON, validation, logging |
stdio |
| StoryForge-Agent MCP | research_topic() · create_video_script() — Gemini + Tavily |
stdio via FastMCP |
| Job-Recommendation MCP | fetchlinkedin() · fetchnaukri() — live job matching pipeline |
stdio via FastMCP |
| Project | Architecture | Stack |
|---|---|---|
| AI-Infrastructure-Copilot | LangGraph agent graph · Qdrant RAG · PostgreSQL · Slack · Auth middleware · Rate limiting | LangGraph Qdrant FastAPI Claude |
| Multi-Agent Purchasing System | Multi-agent procurement workflow · AMD Instinct GPUs · Google ADK | Google ADK AMD GPU Python |
| Real-time Job Recommendation | GPT-4o resume analysis · live job scraping · MCP server · Streamlit | GPT-4o FastMCP Streamlit Docker |
| StoryForge Agent | Autonomous research + script generation · Gemini · Tavily web search | FastMCP Gemini Tavily |
| ClinInsight Medical AI | Clinical AI assistant · medical NLP · conversational interface | LangChain Python |
| Project | Task | Model | Result |
|---|---|---|---|
| EmotiCare | Emotion detection · crisis support · chatbot | BERT fine-tuned | ⭐ 1 star · crisis_detector + chatbot_engine |
| TextSummarizer | Abstractive summarization | BART / T5 fine-tuning | CI/CD · Docker · src layout |
| Topic Modeling | LSA · NMF · LDA · coherence tuning | Gensim · spaCy · pyLDAvis | CI (3 Python versions) · pre-commit |
| NLP Political Leaning | Political bias detection in text | TF-IDF · Naive Bayes · SVM | Multi-class classification |
| Sentiment Analysis | Multi-class sentiment pipeline | Transformers · VADER | Full NLP pipeline |
| Fraud Detection | Imbalanced fraud detection (0.17% fraud rate) | XGBoost · CatBoost · SMOTE | 86% recall |
| Traffic Forecasting | Time-series highway volume forecasting | LSTM · TensorFlow | Sequence modeling |
AI Infrastructure & LLMOps
Platform & Orchestration
Observability
Agentic AI & MCP
ML & NLP
CI/CD & Security
ML Platform & LLMOps Engineer — I close the gap between a working model and one serving production traffic reliably.
Most teams can train a model. Few can serve it at scale. I build the layer in between:
- GPU-aware Kubernetes scheduling — right model on right hardware, production-grade Helm charts
- Inference-specific autoscaling — HPA on queue depth (
vllm:num_requests_waiting), not CPU - Full-stack observability — TTFT/TBT/E2E in every response, 12 Alertmanager rules, OpenTelemetry traces
- Agentic systems — LangGraph workflows, Qdrant RAG, MCP servers for Copilot Studio and Claude Desktop
- NLP pipelines — BERT fine-tuning, BART/T5 summarization, topic modeling with production CI/CD
Open to full-time roles: MLOps · ML Platform Engineering · LLMOps · AI Infrastructure · Applied AI · Founding Engineer (AI)
Sunnyvale, CA · No sponsorship needed · Available now [email protected]
MS Data Science · University of San Diego · GPA 3.9 · Dec 2025 · Built with kubectl, pytest, and too much Prometheus