A Mexican Electronics Engineer and Software Engineer (Backend, Cloud Infrastructure & AI Agents) based in a rainy and techie city, Seattle, WA.
I bring the heat from a rainy city, building smart solutions from the hardware up to the cloud.
I am an innovator who loves a good challenge, whether building innovative tech or tracking down a rare Pokémon card. My passion is combining cloud infrastructure, AI agents, and hardware to build smart and efficient systems that get the job done faster 😎🤖🦾.
- 🔭 I'm currently exploring new opportunities as a Software Engineer focused on backend, cloud infrastructure, or AI agents. I have nearly 4 years of industry experience across Mexico and the United States:
- Microsoft (2022–2025): built Python Hardware-in-the-Loop validation pipelines for Azure servers (~10,000 machines), containerized telemetry services, and Azure CI/CD automation that cut failure rates from ~30% to ~10%.
- Globant (2021–2022): designed middleware and REST APIs for 8 client projects across AWS, GCP, and Azure, including a FastAPI booking app serving 1,000–2,000 daily users.
- 🤖 I'm currently building LLM agents and self-hosted model serving, plus reinforcement learning experiments (DQN, A2C, PPO).
- 🌱 I'm also taking an iOS App Development course.
- 😄 Interests: learning new things (everything! Including Music Theory and Cooking).
- 😄 Pronouns: He/Him 💁🏻♂️
- ⚡ Fun facts:
- I love matcha.
- I am also a Bachata and Salsa dancer! 🕺🏻
- I love to spoil my Wife (and poke her 👀)
Email: joslo2345contact@gmail.com
LinkedIn: Jose Lopez
Tips: Buy me a coffee on Ko-fi ☕
An AI agent that detects GPU faults in fleet telemetry, diagnoses each one with cited repair runbooks, and holds risky repairs for human approval.
- Streaming telemetry pipeline (FastAPI, Kafka, TimescaleDB) replaying a real Alibaba GPU cluster trace across 64 GPUs. The anomaly detector catches 99% of injected faults at 100% precision.
- Tool-calling LLM agent (Qwen3 via Ollama, MCP) with hybrid RAG search (pgvector + PostgreSQL full-text). A held-out eval harness raised diagnosis accuracy from 76% to 88% and fix accuracy from 60% to 92%, at $0 model cost per incident.
- React/TypeScript web app and Slack bot with role-based approvals and an audit log, deployed with Docker, Helm on Kubernetes, Terraform (Azure AKS), and GitHub Actions CI with security scanning.
- Served an open LLM with vLLM on Kubernetes (Helm, Terraform). Switching to it was config-only, and it got 23/25 diagnoses right.
- Autoscaled on queue depth (KEDA, Prometheus, Grafana). The hosted fallback finished all 4 of 4 kill tests.
- Load-tested with replayed agent traffic: prefix caching doubled throughput, and break-even vs. a hosted API is ~170 incidents/day.