Eitan Rosenzvaig

I’m a Staff-level Machine Learning Engineer with 15+ years shipping production ML and LLM systems that move business metrics—from recommendation engines and churn prediction to real-time scoring pipelines processing millions of events and agentic LLM workflows in production.

At Paxful, I was the first ML hire with an open scope. I led a structured roadmap-discovery process to pick the highest-ROI ML bet, then designed and shipped a Kafka-based real-time scoring pipeline (~1M events/month peak ~20 events/minute) using XGBoost served from a Python/Flask microservice. I paired the model with a two-tier action policy (anti-jitter logic + human-in-the-loop review) to align performance with moderator capacity—quadrupling daily fraud catch while shrinking human review from 10 to 2 FTE.

I also built a production RAG case-review system for moderator ban/no-ban decisions, using a tool-calling LLM agent to retrieve account history, trade records, and fraud-graph context and to generate factual, auditable case briefs. To harden the LLM layer, I added an evaluation harness (LLM-as-judge scoring with human spot-checks, prompt versioning, and regression tests gating every prompt change).

Earlier, at Machinio I owned marketing-and-product ML end-to-end: built the company’s first recommendation engine (TensorFlow on AWS), shipped churn and lifecycle messaging, and created a personalized email send-time system (+18% open rate, +23% email-attributed revenue). I integrated LLMs (GPT) for large-scale product categorization with embedding/vector-search grounding (RAG-style), using confidence-gated human review and a gold eval set—cutting uncategorized listings from 15% to 2%. I’ve also led platform engineering work (Gatsby to Next.js at Aleph Beta) and co-founded NeuroCam, bringing an edge + central-AWS inference architecture and YOLO-based detection into unattended production across ~30 cameras and multiple municipalities.