Skip to content

About Me

I’m an ML Systems Engineer at Mozilla.ai, working in an experimental role focused on system-level reliability in AI.

My job operates under a simple premise:

AI systems are sociotechnical systems, where the only components more unreliable than the AI itself are the humans interacting with it. Someone needs to manage that risk.

I build infrastructure for making ML systems more resilient under these conditions and work on open-source projects at Mozilla.ai that make model behavior easier to ship, reason about, and constrain.

If you're curious to learn more about how systems engineers fit on an ML team, these slides from a recent workshop summarize what we've learned so far.

Open source

I maintain and contribute to several open-source projects at Mozilla.ai.

Some are direct outputs of the systems-engineering experiment above; others are adjacent infrastructure that became necessary once we started treating failure as a first-class concern.

  • Encoderfile — A single-file distribution format and inference engine for transformer-based encoders, designed to be offline, self-contained, and compliance-forward.
  • Choice-First Stack — Universal, composable interfaces for LLMs, agents, MCP servers, guardrails, and more.
  • cargo-goose — A SemVer bumping tool for Rust projects that treats pre-releases and real-world versioning workflows as first-class, instead of edge cases.

Outside Mozilla.ai:

  • Ptolemy — An open-source ML observability platform grounded in systems engineering principles

Previously (pre-Mozilla.ai)

  • Head of AI, Seek AI (acquired by IBM) — Semantic parsing, text-to-SQL, research → production pipelines.
  • MS, Computational Linguistics — Montclair NLP Lab — Research in lexical semantic change detections for morphologically rich, low-resource languages
  • BA, St. John's College — Philosophy, history of science and mathematics

Writing & collaboration

I occasionally write, teach, and run workshops—usually about:

  • why alignment is a systems problem,
  • how failures propagate in agentic setups,
  • and why benchmarks lie (politely).

Things I Do Offline

I enjoy cooking, travel, and people with caustic senses of humor. I am obnoxiously proud of being verified at Bassiani, and my favorite subway line is the Q. My favorite authors are Evelyn Waugh, Mikhail Bulgakov, and Giuseppe di Lampedusa.

Contact

Contact me at raz [at] besaleli [dot] io if you are working on something weird, beautiful, or both.