Location: Tampa Bay Area, Florida, USA

  Remote: Yes, remote only. Async-preferred.

  Willing to relocate: No

  Technologies: Rust, Python, C++/CUDA, C#/.NET, TypeScript, Go, Kotlin, SQL | PyTorch, Hugging Face, Burn, ONNX Runtime, llama.cpp, LiteRT-LM, Whisper | SQLite, Docker, Linux, Cloudflare

  Résumé/CV: https://github.com/IsaacLevinsky/resume

  Email: ilevinsky36@gmail.com

  I train small language models from scratch and ship them into production myself - on a single 8-year-old Quadro GP100 workstation, no cloud spend.

  - RustMind: a complete LLM training stack written from scratch in Rust on Burn - model, byte-level BPE tokenizer, corpus pipeline, GPU training. Ladder from 21M to 125M; the 21M encoder is trained to production quality and ships in
  software today. Public pipeline proof (MIT): https://github.com/IsaacLevinsky/rustmind-mini-embed-base

  - Root-caused a generation-collapse failure to non-stratified validation splits and over-memorized templated data, then rebuilt the corpus pipeline around document-aware chunking.

  - LookingGlass: zero-dependency Rust workspace-analysis engine. 3,441 files in ~47ms. Single binary, no Python or Node runtime.

  - Orion: C++/CUDA data engine. 1.15M rows in 1.276s (~901K rows/sec).

  - 17 offline-first apps published across Google Play, Amazon, and Samsung - native C#/.NET Android, on-device Whisper and ML Kit, no ads or accounts: https://play.google.com/store/apps/developer?id=MCMLV1+LLC

  - Live AI in production on three surfaces, including a chatbot on mcmlv1.com running on my own hardware.

  Give me a constraint, a spec, or just a problem, and I design and deliver the whole system end to end. Full-time or contract.