Rafael Levi is a developer-relations professional and web-data specialist at Bright Data who helps AI agents retrieve reliable, current public information. He focuses on a foundational weakness of automated research: agents can produce convincing answers even when blocked requests, CAPTCHAs, empty responses, or outdated data prevent them from accessing the web.
In early 2024, Levi was a senior solutions consultant at Bright Data, advising developers on proxy selection, browser automation, infrastructure, and scraping reliability. Later that year, as a senior solution architect, he tackled large-scale travel-data collection, including rapidly changing fares, anti-bot defenses, cloud-hosted browsers, and parallel browsing sessions. He subsequently brought that infrastructure experience into developer relations and AI-agent workflows.
Silent web-access failures: Agents may substitute stale training data or fabricate plausible answers when a public page is blocked or unavailable. Levi demonstrates these failure modes across property, professional-networking, social-media, and shopping websites, arguing that trustworthy agents must distinguish successful retrieval from apparent access.
Reusable agent-generated scrapers: Levi uses Claude Code, Bright Data's public-web MCP server, and its agent skills repository to inspect websites, identify selectors, and generate conventional extraction scripts. Running those scripts directly avoids repeatedly processing entire pages through a language model and produces structured data that can be scheduled and validated. These repositories belong to Bright Data; Levi demonstrates their application without claiming personal authorship.
Self-healing data pipelines: His production-oriented extraction workflows monitor recurring collection jobs, flag missing fields or changed website structures, and use agents to repair broken scrapers. Applications range from commercial research to alerts for public housing listings.
Public-data boundaries and focused tools: Levi limits these workflows to publicly accessible information, distinguishes it from material behind authenticated accounts, and advises checking website terms. He also favors selectively enabling relevant MCP tools, preventing unnecessary capabilities from consuming an agent's context.
More recently, Levi has applied the same retrieval concerns to video discovery for agentic world models, identifying useful scenes, actions, objects, and environments in public video for training-oriented workflows.
A live headphones search shows how an agent can inspect a protected website, generate a reusable scraper, and reserve model work for building, checking, and repairing the pipeline.
Ten thousand products should not mean ten thousand LLM extractions
Reliable web agents need more than plausible answers: they need usable retrieval tools, explicit failure reporting, and a clear distinction between accessing a page and verifying its contents.
▶ Watch ↗