Surabhi Trivedi

Designed and built an evaluation harness for a bilingual production RAG assistant, including a 59-item golden test suite and LLM-as-judge scoring pipeline. Implemented retrieval-quality metrics and agentic adversarial-probe evaluation loops to surface novel failure modes.

Owned the Good Merchant Gone Bad fraud-risk portfolio end-to-end, managing ~$1.5M in average monthly losses across ~200M daily transactions. Reduced monthly losses to ~$600K and improved chargeback/loss recall by ~15% through real-time heuristics, experimentation, and ML models.

Developed a text-classification system using GPT-3.5 to classify IRS documents, achieving 87.5% recall and reducing client work-hours by over 80%. Built a custom spaCy NER model and a CNN-LSTM image-captioning system for law-library photos.

Led a cross-functional team developing a credit-scoring model for a personal-loan portfolio with 500K+ annual applications, supporting 5x portfolio book growth while reducing default rates by 41%. Designed segmentation, propensity, and pricing models for installment-loans portfolios, contributing an estimated €0.8M annual impact.

Built an ML pipeline for the Chilean Department of Labor to predict high-risk facilities and optimize inspector allocation, increasing violation detection from 36% to 68% and helping protect approximately 9,800 workers.

Developed predictive regression models to forecast load, optimize manpower allocation, and reduce turnaround time.

Pursued a Master of Science in Data Science, focusing on applied machine learning and data-driven decision making.

Completed a Postgraduate Diploma in Liberal Studies through the Young India Fellowship, emphasizing interdisciplinary learning and leadership.

Earned a Bachelor of Technology in Biotechnology, gaining a strong foundation in scientific principles and analytical skills.