Kevin Benassuly

• o Designed end-to-end hybrid ML/LLM automated grading pipeline for large universities — combining classical ML for high-confidence cases and LLMs for ambiguous ones, achieving 60% auto-grading coverage at 95%+ accuracy across cohorts of 1,000+ students and reducing manual teacher workload by 60%, providing students fairer and faster feedback.

• o Built continuous quality loop (LLM explainability, grader bias detection, rubric flagging) that eliminated 50% of model errors by surfacing grader inconsistencies and unlearnable rubric categories.

• o Partnered with product owners and IT domain experts to define a full IT support ticket taxonomy (5 macro categories, 54 subcategories) — shipping a two-stage BERT cascade classifier with 70% auto-routing coverage at 93% accuracy and 80%+ macro F1.

◦ o Designed measurement framework to evaluate classifier performance in production: tracked auto-routing rate, per-category precision/recall, and expert confirmation rate via thumbs-up feedback mechanism (95% confirmation on auto-routed tickets).

◦ o Reduced IT expert workload by ~30% by combining automated ticket routing with automated resolution of repetitive high-volume tasks (password resets, login issues, onboarding queries).

• o Built BERT + Elasticsearch semantic search system powering the Market Insights product — enabling analysts to search 10,000+ financial documents, achieved a MAP of 0.76 and a p@5 of 0.85.

• o Trained Named Entity Recognizer from scratch with no pre-labeled dataset to extract 10 financial entity types from corporate annual reports and earnings – achieving 83% entity-level F1 score.