I'm Arun Addagatla, an AI systems engineer who builds infrastructure for durable, production-grade agents - not slide-deck demos. As the first engineering hire at Lamatic.ai (TechCrunch Startup Battlefield 200, Cloudflare Workers Launchpad), I've shipped over 80% of the platform: serverless agent execution, multi-agent orchestration, production memory and RAG layers, and the runtime glue that keeps workflows alive past a single HTTP request.

At Lamatic, I architected a serverless executor handling Millions of Monthly agent runs and a deployment engine that cut latency from 2 minutes to 15 seconds (~87%). I built an LLM-as-a-judge evaluation framework, a Multi-Agent Supervisor with semantic RAG, Kubernetes ETL pipelines (Drive, S3, SharePoint), and a multimodal hiring agent that reduced recruiter workload by 70%. Most of the hard lessons came from reliability at scale—retries, per-tenant isolation, queue orchestration, and what breaks after the millionth run.

Before Lamatic, at Samespace, I fine-tuned Mistral and LLaMA-2 with LoRA/PEFT, built an LLM inference engine at 106 tokens/sec (dynamic batching, multi-GPU), and drove Whisper V3 ASR to 0.1–0.4s latency with ONNX, TensorRT, and Triton—that's where I learned to make AI dependable under load, not just impressive in a notebook.

Outside of work, I write on Medium about agent infrastructure, inference optimization, and MCP—70k+ reads from engineers who want the hard systems problems explained clearly. I hold a B.E. in Computer Engineering from the University of Mumbai (9.6 GPA).