An N.

I build AI systems that run in production - not demos. 5+ years taking models from research to real-time services that hold up under load: computer vision on GPUs and edge devices, and LLM chatbots, voicebots, agents, and RAG platforms serving enterprise customers. What I do ✅ Computer Vision & Deep Learning - pose estimation, face recognition, video analytics. Training, fine-tuning, and shipping custom models (PyTorch, MMPose/MMDetection, YOLO, ONNX) ✅ Real-time C++ inference services - gRPC streaming, H.264 decode pipelines, thread-pool model serving, multi-model orchestration on a single GPU ✅ NVIDIA GPU optimization - TensorRT engine building, fp16/fp32 trade-offs, CUDA memory tuning, and the profiling that finds where the time actually goes. Recent example: cut p95 latency ~4x on a T4-based service by finding the bottleneck was CPU scheduling, not GPU compute ✅ Chatbots & voicebots - production multi-tenant conversational AI platform, serving multiple enterprise customers. Chat and voice assistants with intent detection, dialogue state tracking, multi-turn memory, and a graph-based conversation state machine. Real-time streaming for both chat and voice, guardrails (NeMo Guardrails), and automated conversation quality evaluation (fluency, coherence, empathy, factual accuracy) ✅ LLM agents & MCP - multi-agent and deep-agent systems (LangGraph, LangChain, OpenAI SDK), plan-then-execute orchestration with human-in-the-loop approvals. Built MCP servers (FastMCP) and scaled agents over a 170+ tool MCP catalog with dynamic tool loading, cutting prompt tokens ~5x per run ✅ RAG & search - production RAG pipelines: document parsing, chunking, embeddings, hybrid retrieval (dense + BM25) and reranking on Qdrant, Elasticsearch, FAISS, Pinecone, Cohere. Query rewriting and evidence grounding to reduce hallucination ✅ LLM evaluation & observability - Langfuse tracing, offline eval harnesses, LLM-as-a-judge, regression evals, per-run token cost tracking ✅ LLM backends - FastAPI, async Python, WebSocket/SSE streaming, Redis semantic caching, MongoDB. OpenAI, Azure OpenAI, Ollama, open-source LLMs; fine-tuning LLMs and embedding models ✅ Cloud & MLOps on AWS - EC2/ASG, ECR, ElastiCache, RDS, S3, ALB/NLB, IAM, VPC, Terraform for all of it. Immutable release tagging, GitHub Actions CI/CD with OIDC, zero-downtime rolling deploys, CloudWatch and Prometheus. Docker, Jenkins ✅ Reproducible ML pipelines - dataset manifests with content hashing, artifact custody in S3, provenance so any shipped model traces back to its exact weights, config, and data ✅ Edge AI - Jetson Nano/Orin, Raspberry Pi; model conversion, quantization, native on-device inference ✅ Mobile - CoreML and TFLite conversion, native inference on iOS and Android How I work - I measure before I optimize. Most "GPU problems" turn out to be somewhere else entirely - Clean, tested C++ and Python - with the test that actually fails when the code breaks - Infrastructure as code, so a deploy is repeatable rather than remembered - Direct communication: if an approach won't work, you'll hear it early, with the reason I've worked on self-driving cars, face recognition, real-time posture analysis, a multi-tenant conversational AI platform, and an enterprise agent platform. What connects them is the same thing: making AI work reliably outside a notebook. 🚀 Let's turn your idea into something that runs in production.