NATHAN ORDONEZ

Aspiring Deep Learning Research Scientist
generalist, deep-working, entrepreneurial
fascinated by information representation
& neural-based methods

Professional Experience

  • Research Associate @ IBM Research - Zurich (2025) [9mo]
    • Implemented independent KV-caches (block-sparse attention) in vLLM, achieving:
    • • improvements in KV cache hit rates
    • • long-context cold-start prefill speedups
    • • detailed evaluations / output quality study
    • • Explored state combination methods for all hybrid LLM architectures
    • (SSM, Liquid, LightningAttn.)
    • • Supervised by early and major vLLM contributor
    • └> Resulted in public release as IBM's official vLLM public fork
    • └> + Arguably the fastest end-to-end implementation of LLM inference
            with this feature among the labs
  • Project & Design Lead @ Hololearn TU Delft research group (2022) [5mo]
    • Based on EU-level research grant
    • Assembled and led a 5-person team to develop and
      design hologram-based virtual teaching system
    • Involved full-stack development including 3D+LiDAR data processing
    • Attracted collaboration from Technologico de Monterrey
    • └> led to further development + CEUR Workshop publication (6 citations)
  • Founding Data Scientist @ CorTexter (2020-2021) [1y]
    • Developed+designed core ML pipeline for advanced
      NLP-based (neural + linguistic) job vacancy analysis
    • Built multilingual neural job title search engine
    • Trained emoji-prediction language model
    • └> pipeline still in use after leaving
  • Startup Analyst @ Plug&Play Tech Center (2020-2021) [9mo]
    • Major California-based VC fund
    • Analyzed hundreds of startup pitch decks
    • └> Praised for story-driven final pitch analysis

Research Highlights

  • Master's Thesis (2024):
      Introduced novel DL architectures for DNA basecalling
      (up to 85% throughput improvement over best Oxford Nanopore model)
      Introduction of Chain-of-thought technique to basecalling
      └> led to continuing partnership with Tenstorrent (AI accelerator company)
  • Elsevier Publication (2024) (33 citations):
      GNN-based power network prediction model
      achieving 48x speedup over non Deep-Learning-based method
      └> led to further work in the area by co-authors
  • Arxiv Preprint (2023):
      Pioneered RL loss allowing an agent to navigate
      through temporary blindness (open loop control)
      └> led to a successful research grant

Achievements & Projects

    • - Ready to Startup jury winner for space transport platform
      (validated by NASA contractor Axiom Space)
    • - 1st place Yes!Delft Students hackathon for coral reef startup
    • - 3rd place Best Delft ideation contest
    • - Selected Nova Network member (top ~3% of applicants)
    • - Host of UvA/Slim Radio Monday Morning Podcast
      interviewing high-achieving students