Work

Download resume
  • May 2023 - Present
    Apporto
    Machine Learning Contractor
    • Served as the primary ML engineer for multiple education-technology products, translating high- level product goals into model designs, algorithmic specifications, and implementation plans.
    • Designed an academic-integrity detection system using chunk-level probabilistic analysis of writing style, behavioral production history, and document provenance signals.
    • Created a submission ordering algorithm to minimize repeated failure modes due to delayed in- struction updates, based on an initialization heuristic and greedy optimization.
    • Designed and optimized a fuzzy text-search algorithm for locating small-to-medium text blocks within large documents using a two-stage bounded search procedure.
    • Contributed to core architecture design necessary for concurrent, high-throughput performance of the grading products.
  • Feb 2022 - Jan 2023
    Branch
    Technical Lead
    • Led backend and DevOps engineers building an NLP filtering product based on a few-shot text classification method I developed.
    • Architected distributed microservices for high-volume text ingestion, filtering, and output delivery.
    • Met regularly with developers to perform code reviews and ensure their alignment with project goals and timelines.
  • Feb 2022 - Jul 2022
    Opsis
    Machine Learning Contractor
    • Conducted feasibility research on neural approaches to point-cloud volume estimation for a visual calorie prediction system.
    • Implemented literature baselines, custom PyTorch layers, and experimental model variants to compare architecture choices under limited research data.
    • Summarized experimental results and recommended next technical directions to the principal investigator.
  • Jul 2021 - Feb 2022
    Branch
    Machine Learning Specialist
    • Joined as the sole ML engineer, leading R&D to assess the feasibility of an unsupervised approach for filtering high-throughput text streams.
    • Built a few-shot filtering method based on regularized embedding spaces and an SVM-style clas- sifier, requiring only 5–10 examples and achieving ∼0.85–0.95 F1, depending on task.
    • Developed custom RoBERTa-based embedders with higher-dimensional embeddings and latent- space regularization to improve separability and generalization with sparse labels.