Guadalupe Higuera

• Core engineer from inception on a production agentic-AI platform serving 10,000+ users across 6 business lines, including policy, finance, and investments
• Built orchestration APIs routing thousands of daily chats across Claude, Gemini, and GPT models; allowing for model agnostic agent orchestrations allowing users to switch models based on both performance and costs
• Designed and built the configuration APIs for users to create, modify, and deploy their own AI agents
• Developed an MCP server allowing for agents to search SharePoint, Confluence, GitHub and SQL databases; support for multi agent summarization allowing for agents to search relevant data much quicker, speeding up agent orchestration timing
• Own deployment and availability on a weekly release cadence; built self-healing infrastructure on Kubernetes (OpenShift); sustaining 99.99% availability across an enterprise environment
• Mentored junior engineers and interns on both AI engineering and software engineering best practices

• Shipped production .NET services and React front-ends for internal and customer-facing applications

iRacing AI Race Strategist | GitHub | Demo

• Fine-tuned Llama 3.1 8B with QLoRA (PyTorch, Hugging Face) on 9,269 category-balanced synthetic examples using Claude API
• Improved evaluation scores and eliminated hallucinations, measured by a 12 metric weighted scorer along with a blind LLM-as-judge A/B comparison
• Designed event-driven LLM orchestrations across a variety of event racing scenarios for a real-time voice race engineer with 800 – 1,100 ms telemetry to speech latency

TestPilot — Self-Repairing Agentic QA | GitHub

• Built an agentic QA tool (TypeScript, Playwright, OpenAI) that turns plain-English stories into generated Playwright tests, and auto repairs Playwright tests on safe UI drift
• Designed layered AI guardrails: veto-only vision diagnosis, repair guards that preserve test intent and a human approval PR gate; ingests GitHub/Jira stories as an MCP client

Titanic Historical RAG — Contradiction Detection | GitHub | Live

• Built a RAG system using over 3,400+ pages of testimony (40K chunks, 158 witnesses) using OpenAI embeddings, Pinecone and Q&A boundary chunking to surface contradictions between witnesses
• Implemented LLM-as-judge contradiction detection (Claude Haiku 4.5) with structured JSON output, confidence scoring, and a DynamoDB verdict cache