Full-stack AI engineer. I ship production agentic systems end to end: voice agents, LLM pipelines, evaluation and observability, and the data and safety layers around them.

I build AI that operates in the real world and gets judged by outcomes, not demos. Self-taught and operator-first: built and sold an e-commerce business ($4.6M, private-equity exit), ran agency-side marketing automation, and now build production AI systems independently, deployed into real client operations. My entire practice is AI-native: Claude Code, Cursor, and Codex across the full development cycle, every system measured by a revenue or efficiency metric.

Voice-AI and automation platform: multi-tenant platform for white-labeled AI voice agents. Per-tenant Postgres row-level security, Stripe prepaid wallets with atomic billing that stays consistent under concurrent webhooks, and a two-stage call-classification pipeline (deterministic rules plus an LLM tie-breaker) that scored 14/14 on its eval set separating live humans from call-screening bots.

AI-native healthcare operations: ambient clinical scribe with real-time transcription and structured notes, PHI-first architecture with zero-trust access and audit trails, and a patient-facing voice agent.

Autonomous pipelines: systems that ingest, score, and act on data on a schedule (market monitoring, lead generation, business intelligence), each with run-cost budgets treated as first-class design constraints.

• Agents should own end-to-end workflows, not steps. If it can't own an outcome, it's not ready.
• LLM-as-judge everywhere, but the threshold lives in code: testable, versioned, immune to model mood.
• Build systems that absorb capability improvements rather than get replaced by them.

dev@erickpaniagua.com · erickpaniagua.com · github.com/erickcxc · Detroit, remote (US)