Hire Subham J. - RAG Engineer

Benevolent Stereotyping Analysis Architected a specialized NLP pipeline in Python to automate the detection of benevolent stereotypes by distinguishing whether phrases were positive, benevolent, hostile, and irrelevant. Engineered a few-shot classification model using the SetFit (Sentence Transformer) framework, moving beyond traditional zero-shot limitations to capture semantic nuances regarding "agency" and "natural ability." Propelled prediction accuracy from 55% baseline achieved by a zero-shot approach to 86.8% by implementing a contrastive learning strategy, training the model to physically separate "body-relevant" attributions from "mind-relevant" skills in vector space. Resolved critical semantic ambiguities specifically between benevolent stereotypes and positive compliments by utilizing a targeted contrastive dataset to force the model to disentangle natural traits from learned ones

* Leadership Framing in Media * Developed a robust, automated data pipeline in Python to parse, preprocess, and structure thousands of PDF news articles from reputable organizations. * Engineered a zero-shot classification model leveraging the powerful DeBERTa-v3-large transformer architecture to categorize nuanced linguistic frames within sentences. * Augmented the model's prediction accuracy by over 30% by implementing innovative, prompt-driven hypothesis templates. * Managed the entire project lifecycle, from raw data ingestion and batch processing of large text volumes to producing a final, labeled dataset with tens of thousands of predictions.