I’m an AI engineer from Santa Tecla, El Salvador. At Rivka Development, a US consulting firm, I build AI agents and RAG systems for client projects and take them to production, with evaluation and per-query cost tracking built in.
I also lead product and engineering at Eclalune, my own beauty and personal care company in El Salvador.
Before that I was a full stack engineer at Veronica’s Insurance, building AI tools for underwriting and sales. Two years ago I was doing Tier-2 technical support; I got here by building things that had to work.
Most of my client work is under NDA. I’m happy to walk you through the architecture on a call.
I build AI agents and RAG systems for client projects in Python and TypeScript, mostly FastAPI and Hono on Cloudflare Workers and Azure, with PostgreSQL and pgvector. I also add evaluation and per-query cost tracking to our RAG systems, and ship anomaly detection and similarity search into production.
I built AI tools for teams like underwriting and sales, with retrieval and prompt orchestration for claims and customer service. React on the frontend; AWS Lambda, DynamoDB and S3 on the backend, tuned for cold starts and cost.
I helped the company adopt generative AI with Python automations and self-hosted tools on Linux, and trained non-technical staff to use LLMs in their daily work.
A data-driven beauty and personal care company for Latin America, selling directly to consumers in El Salvador. I lead product and engineering, and I built the web app it runs on: the storefront, with catalog, cart and WhatsApp checkout, and the back office for orders, purchasing and inventory with lot and expiry tracking.
In 2026 I went to the Plan B Summer School in Lugano, Switzerland, with El Salvador’s delegation, on a Plan B Network scholarship.
I built honeybadger-ai, a self-hosted tool that sets limits on what AI agents can do with a Bitcoin wallet. It never holds the keys.
Proof of concept for Tether
A requirement of the Plan B program: predictive maintenance for Bitcoin mining fleets. Isolation Forest, an LSTM autoencoder and XGBoost flag hardware failures; an LLM assistant on top reads only the model output. Trained on synthetic fleet data.
Claude Code and Codex are part of my daily workflow. I build MCP servers that connect agents to APIs, databases and internal tools, and I review every change before it merges.